Open Source LLMs

As of August 2026, AIDiveForge tracks 125 open source llms. The top three by verified-data score are Nimbus, Open-Kritt, and Xalgorix. Curated open source llms tracked by AIDiveForge. Each project has a verified public source repository. Listings are verified against each tool's live website and re-checked regularly.

Last updated July 29, 2026 · 125 tools

Ranked by AIDiveForge's verified-data score: data completeness, verification recency, community rating, and real visitor engagement. How we rank · No tool can pay for placement.

1. 

   1. Nimbus

   Nimbus runs a ReAct planning loop that maps a natural-language request to actual cloud actions: querying live AWS or GCP telemetry, generating infrastructure changes, opening PRs on connected repositories, and updating shared architecture diagrams. Approval gates sit between the agent's plan and execution, so nothing ships without a human sign-off. That model works well for incident diagnosis and routine cost optimizations. Where it strains is on cross-account, deeply custom IAM environments — the agent's tool set reflects the scaffolding its maintainers have wired up, and anything outside that surface area requires you to extend it yourself. Self-hosting via Docker or source install keeps sensitive cloud credentials off third-party infrastructure, which is the primary reason platform teams choose it over a SaaS alternative.

   PaidOpen SourceAPISelf-hostedVerified Jul 8, 2026
2. 

   2. Open-Kritt

   The tool runs parallel AI agents across a codebase, so vulnerability discovery that would serialize into hours on a single-context scan distributes across concurrent analysis threads. It targets security researchers and bug bounty teams who need to sweep repositories at scale, not review a function at a time. Self-hosting is supported under AGPL-3.0, which means your code and findings never leave your infrastructure — a requirement for any org with compliance constraints. The open-source core is inspectable and forkable, but managed scans are a paid-only feature, so teams that want the hosted workflow face a significant spend threshold. The page describes GitHub integration as a first-class path, making it a practical fit for teams already running security workflows inside existing CI infrastructure.

   PaidOpen SourceSelf-hostedVerified Jul 21, 2026
3. 

   3. Xalgorix

   The core loop is detect, chain, verify: the agent runs reconnaissance through injection through authentication testing, then executes a dedicated validation phase before anything reaches your report. On a public deliberately-vulnerable target, the vendor documents 9 verified findings including a CVSS 9.8 RCE in 17 minutes. The REST API and cron-style scheduling let security teams wire scans directly into CI/CD gates, so releases block on verified findings rather than scanner noise. Where the architecture shows its limits: scan depth and concurrency are credit-gated, and teams running continuous coverage across a wide attack surface will need to budget credits carefully. Self-hosted deployment is listed as an option for teams with data-residency requirements.

   PaidOpen Sourcefrom $1 per scanAPISelf-hostedVerified Jul 8, 2026
4. 

   4. Emem

   emem stores facts as short, signed tokens — each one a content-addressed handle that any agent can carry through a summarization pass, hand to another agent on a different model or vendor, and resolve back to the exact signed bytes without trusting whoever sent them. The verify step is offline: recompute the hash and ed25519 signature yourself, no server call required. Cold resolution runs around 180 ms; warm cache hits around 10 ms, with every receipt reporting its own latency stats. The honest caveat from the vendor's own benchmarks: against a bare inline number, a single emem token costs 5.8x more context — the savings only appear when you bundle multiple facts into one round trip.

   PaidOpen SourceAPIVerified Jul 23, 2026
5. 

   5. Genesys

   Genesys stores what you share in a causal graph you own, then surfaces that context to any app that speaks MCP — so Claude already knows what you told ChatGPT, without you repeating yourself. The graph explains its own reasoning: ask why it remembers something and you get the actual chain of connections, not a confidence score with nothing behind it. Memories fade by a scoring formula tied to relevance and reactivation, so stale data drops out without silently deleting things that still matter. The free tier caps writes at 300 stores per month — heavy users or teams running MCP agents hit that ceiling, then face a choice.

   PaidOpen Source$0-$8/moAPISelf-hostedVerified Jul 22, 2026
6. 

   6. Jaybase

   Jaybase stores every agent-generated fact as an immutable, time-stamped record, which means the full sequence of what an agent wrote, when, and why is always recoverable. The vendor describes it as designed for accounting, compliance, and approval workflows where you cannot afford to lose the paper trail. Because it is append-only, there is no overwrite risk — replaying a sequence from any point is a native operation. The library is self-hostable and open-source under AGPL, so it runs inside your own infrastructure without a call home. The project has a small contributor footprint, which means production teams should expect to own gaps in documentation rather than wait for the maintainer to fill them.

   FreeOpen SourceAPISelf-hostedVerified Jul 23, 2026
7. 

   7. Shepherd

   SHEPHERD is a Python substrate from Stanford and Northeastern that turns an agent's execution into a Git-like, reversible trace — so a supervising meta-agent can observe, intercept, fork, and revert any step without rebuilding that capability from scratch each time. The vendor-published benchmark numbers are specific: a supervisor meta-agent lifted pair-coding pass rate from 28.8% to 54.7% on CooperBench; a counterfactual repair meta-agent beat MetaHarness on Terminal-Bench 2.0 by 12.8% while cutting wall-clock time by 58%. The framework is research-grade and open-source, installed via pip. Teams outside the specific use cases the paper targets — runtime intervention, counterfactual optimization, and agentic RL training — will find precious little guidance on how far the substrate stretches.

   FreeOpen SourceSelf-hostedVerified Jul 13, 2026
8. 

   8. firstmate

   firstmate puts a single orchestrating agent — the 'first mate' — in front of you, while it spawns a crew of autonomous coding agents behind the scenes, each isolated in its own git worktree. You describe what needs doing; the crew splits the work in parallel and keeps collisions out of your main branch. The visible session backend means you can watch what each agent is doing without switching tabs. The architecture works cleanly for investigation tasks, parallel fixes, or supervised PR generation — the constraint is that there is no API surface, so anything requiring programmatic integration into an existing CI pipeline has to wire around the tool manually.

   FreeOpen SourceSelf-hostedVerified Jul 14, 2026
9. 

   9. AIfunc

   The tool treats AI calls the way you already treat HTTP requests: stateless, typed, testable, and wired into your existing code with standard language control flow. No canvas, no orchestration runtime, no new mental model. The vendor states the target is the 80% of real-world AI work that is text-in, structured-data-out — sentiment analysis, summarization, classification. Multi-step workflows are composed with the same if-else and loops you already write. Where this breaks: anything requiring memory across turns, autonomous planning, or tool-use loops is outside the design scope entirely.

   FreeOpen SourceAPISelf-hostedVerified Jul 9, 2026
10. 

    10. ami

    Ami builds a context graph in SQLite and markdown on disk — tracking entities, relationships, your past decisions, and your writing style — so the agent gets less hand-holdy the more you use it. It maintains a live to-do list and executes recurring busy work by learning how you handle tasks, not by following a static playbook. The self-hosting story is real: no data leaves your machine, no org-level sharing. Where it breaks is scale and surface area — the repo has 3 commits and 4 stars at time of writing, which means production edge cases are yours to discover and debug, not documented anywhere. Teams that need multi-user workflows, audit logs, or a managed API surface will hit the ceiling fast.

    FreeOpen SourceSelf-hostedVerified Jul 23, 2026
11. 

    11. Bike4Mind

    The workbench routes across 60+ models from OpenAI, Anthropic, Google, and AWS Bedrock through a single interface and API, with a separate lane for open-weight models running on your own hardware via vLLM — the lane no lab can ever sell you or switch off. Sessions, prompts, and artifacts survive mid-conversation model swaps, so when a provider gates its best tier, the switch is a config change, not a rebuild. The agentic layer runs 'Quests' — long-running jobs with a code REPL, search, and MCP access under hard budget caps, so you fire a task and return to results rather than babysitting each step. Where the tool shows its edges: the source-available BSL 1.1 license means self-hosted deployments carry restrictions until the two-year Apache rollover, and teams that need branching multi-agent pipelines beyond single-Quest logic will hit the canvas ceiling fast.

    PaidOpen SourceAPISelf-hostedVerified Jul 8, 2026
12. 

    12. ClawLite

    ClawLite extracts the reliability patterns from OpenClaw and strips the rest to roughly 500 lines of logic. You get lane-based serial execution so tool calls don't interleave, automatic context compaction at 80% capacity so small models don't hit the wall mid-task, and provider fallback so a dropped Ollama instance doesn't kill a pipeline. Skill behavior is configured via markdown files, not code. The ceiling appears fast: there is no API, no web UI, no parallel execution path you can opt into for tasks that actually need it, and the project sits at v0.1.0 — which means the surface area is deliberately small and the community footprint is thin.

    FreeOpen SourceSelf-hostedVerified Jul 24, 2026
13. 

    13. Cognee

    The core workflow is three lines: install via pip, point Cognee at a data source, and your agents start recalling cited facts instead of hallucinating from scratch each session. Graph-structured memory means relationships between entities survive retrieval — not just keyword matches. First-party integrations cover Claude Code, Cursor, LangGraph, and an MCP server, so compatible agents read and write memory without custom glue code. The ceiling appears when your ontology needs get specific: custom data models and permissions controls are available, but tuning graph structure for a niche domain requires real configuration work. Teams that need a drop-in vector store with zero graph overhead typically reach for a simpler solution.

    PaidOpen SourceAPISelf-hostedVerified Jul 13, 2026
14. 

    14. Craft Agents

    Built by the craft.do team for their own internal agent work, Craft Agents connects to external services like Slack or Linear, imports skills from Claude Code, and runs local MCP servers as subprocesses — all without requiring CLI fluency from the people reviewing agent output. The document-based interface lets you run multiple agents in parallel and share sessions, which removes the 'only the dev knows what happened' problem. Where it strains: the GitHub issues tracker shows 311 open issues and 146 open pull requests, signaling a project moving fast with rough edges still exposed. Teams expecting enterprise-grade stability at scale will feel those gaps quickly.

    FreeOpen SourceSelf-hostedVerified Jul 14, 2026
15. 

    15. Cruxible

    Cruxible treats agent-written state as something that must be proposed, reviewed, and evidence-linked before it counts — not just logged after the fact. Every claim carries a receipt, every query can be reproduced, and every write goes through a structured proposal workflow rather than landing directly. The model fits audit-heavy domains where 'the agent said so' is not an acceptable answer. The ceiling appears when you need the tool to decide what to do next: Cruxible is a state layer and workflow runtime, not a planner, so you bring the decision logic yourself. Teams with existing data exports can convert them into structured state, but anything requiring dynamic planning or autonomous tool use sits outside this tool's scope.

    FreeOpen SourceSelf-hostedVerified Jul 9, 2026
16. 

    16. Cybara

    Cybara is a fully open-source, MIT-licensed agent runtime you run entirely on your own hardware, with no required account, no telemetry, and no cloud dependency. The vendor describes 50+ model providers, 80+ built-in tools, and 25+ messaging channels available out of the box — a surface area that takes serious effort to configure correctly before it pays off. The operator controls are the genuine differentiator: per-session tool allowlists, filesystem checkpoints, spend caps, and approval gates mean agents don't act without your sign-off. The breadth is also the trap — teams starting with a single use case will spend time disabling or routing around features that don't apply to them. Where it earns its complexity is multi-channel deployments or on-chain automation where the alternative is building that policy layer from scratch.

    FreeOpen SourceSelf-hostedVerified Jul 17, 2026
17. 

    17. eve

    The platform gives coding agents a native deployment surface — API, CLI, MCP, and agent-callable Skills — so agents ship and iterate on apps without a human relaying commands. Sandboxed VMs let agents run code they generated without that code touching your production environment. Durable Orchestration means a workflow that pauses for minutes or months resumes from the exact checkpoint, not from scratch. The constraint is architectural: there is no self-hosted path, so teams with strict data-residency requirements or air-gapped environments hit a wall before they write a single agent. At that point, the conversation moves to a competitor with an on-premises option.

    PaidOpen Source$20/moAPIVerified Jun 29, 2026
18. 

    18. GSV

    GSV deploys your AI into your own Cloudflare account — not onto a box you manage, but across an edge layer that connects every device you own into one shared context. Your laptop, home server, and phone act as a single computer, and the brain keeps running when none of them are on. Reach it through Telegram, Discord, or a terminal — wherever you already work. The tradeoff is real: GSV requires a Cloudflare Workers Paid account, which means your infrastructure is permanently coupled to Cloudflare until off-platform self-hosting ships. That roadmap item is public, but it is not yet available.

    PaidOpen Source~$5/mo Cloudflare + model costsVerified Jul 2, 2026
19. 

    19. Hearth

    Hearth runs on your own hardware and handles the tasks that usually demand a SaaS subscription: opening applications, reading and writing files, driving a real browser you can watch, and carrying memory of past sessions — all without a single request leaving your network. The MIT license means you can fork it, extend it, and ship modified versions without legal friction. That said, the GitHub repo shows 9 stars and 297 commits from a single-org project, which signals early-stage software rather than a hardened production runtime. Windows is the primary target; Linux and macOS support is not confirmed by the page. Teams that need cross-platform deployment or enterprise support will hit the ceiling fast.

    FreeOpen SourceSelf-hostedVerified Jul 19, 2026
20. 

    20. Isnad

    Isnad attaches provenance metadata to individual claims as they move through agent pipelines, borrowing the narrator-grading logic from classical hadith transmission scholarship to score source reliability at each hop. The vendor describes it as claim-level auditing — you get a trustworthiness grade per claim, not a flat event log. It installs via pip and ships with Docker support and Alembic-managed migrations, which means it slots into existing Python stacks without standing up a separate service. The ceiling appears when your pipeline is not Python-based or when you need a hosted dashboard rather than a library you integrate yourself. Teams outside that boundary are building their own wrapper before they can use the core grading logic.

    FreeOpen SourceAPISelf-hostedVerified Jul 11, 2026
21. 

    21. Kastor

    Kastor lets you define agents, tools, prompts, and models in HCL, validate the spec for broken references and unresolved prompt variables, then compile it to a runnable LangGraph project. The plan/apply/destroy loop catches drift between your spec and what's actually deployed — the same mental model as Terraform, applied one layer earlier. The vendor explicitly labels this an early proof of concept with a working local loop; hosted provider integration is on the roadmap but not shipped. Code generation targets LangGraph only. Teams needing a runtime, a UI, or anything beyond a local CLI will be assembling the rest of the stack themselves.

    FreeOpen SourceSelf-hostedVerified Jul 9, 2026
22. 

    22. Kitaru

    Kitaru wraps your existing agent SDK — PydanticAI, OpenAI Agents, Claude Agent SDK, or raw Python — and turns every model call, tool call, and intermediate step into a durable checkpoint. When you want to ask what would have happened with a cheaper model or a failed retriever, you replay from a specific checkpoint with one override. Nothing re-executes in production. The vendor's own benchmark shows 200 replayed executions on a cheaper model matching outputs in 192 of 200 cases at 84% lower cost. The ceiling appears when your agent's behavior depends on state that Kitaru's adapter doesn't intercept — external side effects or SDK internals the wrapper never sees won't be faithfully replayed.

    FreeOpen SourceAPISelf-hostedVerified Jul 7, 2026
23. 

    23. LangDrift

    Langdrift runs your agent prompts across multiple locales and compares behavior — checking whether tool calls, response structure, and decision paths stay consistent when the input language changes. The core problem it addresses is language-induced behavior drift: the same logical request, rephrased in German or Japanese, producing a different agent output than the English baseline. It fits cleanly into CI pipelines where you need deterministic, repeatable checks across locale variants. The project is built and maintained by a single developer, Rubén González, which means the feature surface reflects a focused scope — not a product roadmap backed by a team.

    FreeOpen SourceSelf-hostedVerified Jul 9, 2026
24. 

    24. Last EHR

    The tool wraps four FHIR operations — patient search, chart read, note creation, and observation recording — into a chat agent that runs as the signed-in user, bounded by whatever Medplum AccessPolicy you already have. Every write surfaces as an approval card showing exactly what will be saved before it touches the chart. The same four tools expose as an MCP server for Claude Desktop or Claude Code, read-only by default. The ceiling is visible early: four tools, one backend, one EHR platform. Teams that need more than that are writing their own.

    FreeOpen SourceAPISelf-hostedVerified Jul 9, 2026
25. 

    25. MemLedger

    The vendor describes MemLedger as a memory framework with an audit trail: every stored fact carries provenance, so when an agent surfaces a stale or wrong preference you can trace the extraction decision that created it. The library includes a policy layer — a `memory.policy.yaml` file — that lets teams quarantine unverified facts before they reach permanent knowledge, which means bad data from one session doesn't silently corrupt the next. An evaluation suite ships alongside the core library, so you can benchmark how well a newer extraction model rebuilds memories from raw history before you migrate. The ceiling appears quickly for teams that need hosted infrastructure, multi-agent coordination, or anything beyond a Python library integration — there is no API, no managed service, and no UI.

    FreeOpen SourceSelf-hostedVerified Jul 10, 2026
26. 

    26. Nova

    Nova runs 24 specialist agents on your own machine against your own API keys, storing everything in local SQLite. The pipeline is explicit — classify, decompose, prepare, then stop at a gate before anything publishes, sends, or spends. That gate is the differentiator. Where it strains: the roster of 24 agents with no hosted fallback means setup requires real technical lift, and the Bun + TypeScript stack narrows who can extend it. Teams that outgrow the built-in connectors or need agents to reason across domains the roster doesn't cover will find themselves writing custom system prompts before they expected to.

    FreeOpen SourceSelf-hostedVerified Jul 26, 2026
27. 

    27. npcpy

    npcpy is a MIT-licensed Python library built around three primitives: Context, Agent (NPC), and Tool — which you compose to wire up single agents or multi-agent teams running against local runtimes like Ollama and llama.cpp or cloud providers. The library's knowledge graph support and multimodal LLM integration live in the same package, so a research prototype doesn't require stitching together three separate dependencies. Where it starts to strain is at the integration surface: documentation is sparse for anything beyond the happy path, and production observability — logging, tracing, failure recovery — is not built in. Teams moving from research prototype to a production deployment will find themselves reaching for additional infrastructure the library does not provide.

    FreeOpen SourceAPISelf-hostedVerified Jul 14, 2026
28. 

    28. Osaurus

    Osaurus runs on Apple Silicon via Ollama, MLX, or LM Studio, fully offline — Wi-Fi off, still working. Drop a folder, assign a task, and agents read, write, and execute against your local file system while you're away. When a task outgrows what the local model can handle, you route to ChatGPT, Claude, or Gemini without losing the shared persistent memory thread. The MIT license means no usage caps and no billing — ever. The hard ceiling is macOS exclusivity: teams on Linux or Windows are looking at a different tool from day one.

    FreeOpen SourceSelf-hostedVerified Jul 14, 2026
29. 

    29. Provena

    Provena wraps around retrieval steps, tools, and context assembly logic to log where every chunk of data came from, hash it for tamper detection, and surface that audit trail when something breaks or an auditor asks. The vendor describes six framework adapters, an MCP server, PostgreSQL storage, and a policy engine — covering most standard Python-based pipelines without requiring a hosted service. Installation is self-hosted and free. The ceiling appears when your compliance requirement goes beyond audit trails: Provena is a passive tracking library, not an enforcement layer, so it records what happened but does not block a bad retrieval from reaching the model. Teams with hard EU AI Act enforcement obligations pair it with a separate policy gate.

    FreeOpen SourceAPISelf-hostedVerified Jul 22, 2026
30. 

    30. Qwen-Image-3.0

    The family spans four distinct problem areas: safety moderation via Qwen3Guard, multilingual translation via Qwen-MT, text-rich image generation and editing via Qwen-Image and Qwen-Image-Edit, and general reasoning via the base Qwen3 models. Self-hosting is a real option — weights are published on Hugging Face and ModelScope, and the Apache-2.0 license means no legal friction for commercial deployment. Qwen-MT's hosted API is a paid-only feature, so teams that want translation without infrastructure management pay for access; everyone else runs inference themselves. The research layer is also public: GSPO, the vendor's proposed fix for RL training instability in large models, is documented and available for teams experimenting with fine-tuning at scale.

    PaidOpen SourceAPISelf-hostedVerified Jul 26, 2026
31. 

    31. Senbonzakura

    The tool identifies the activation-space directions that carry refusal behaviour in open-weight transformer models and edits them out of the weight matrices in a single pass — no gradient descent, no retraining. It extends the Arditi et al. single-direction method by automating direction search (borrowed from Heretic) and then cutting several directions at once, which the author reports moved the needle in practice where single-direction edits did not. The procedure is a one-time weight edit: you run it, you get a modified model file. There is no API, no inference server, and no managed hosting — you run it locally against your own model weights.

    FreeOpen SourceSelf-hostedVerified Jul 19, 2026
32. 

    32. Snippbot

    Snippbot installs via a single pipx command and runs entirely on your own hardware — no cloud relay, no data leaving your network. The core model is a bench of domain-specific agents (bookkeeping, development, design, project management) each with isolated persistent memory backed by a vector store, full-text search, and a typed knowledge graph. Drop those specialists into a shared chat and they pull from their own episodic memory to collaborate on cross-functional work without context bleed between them. The platform is in open beta and the GitHub repository is listed as coming soon, which means community support and third-party integrations are sparse. Teams that need enterprise connectors or a mature plugin ecosystem will hit that wall quickly.

    PaidOpen SourceSelf-hostedVerified Jul 17, 2026
33. 

    33. Talon

    Talon is a self-hosted, MIT-licensed agent harness that runs as a long-lived process with persistent memory, hot-reloadable plugins, and four frontends — Telegram, Discord, Microsoft Teams, and CLI — all sharing one agent core. Swap the backend by changing one line in config.json: Claude SDK, Kilo, OpenCode, Codex, or OpenAI Agents, each implementing the same interface so your plugins and memory survive the switch. Memory is handled through Mempalace — a ChromaDB vector store plus SQLite knowledge graph that retains semantic context across sessions. Background modes (dream and heartbeat) consolidate memory and run proactive maintenance while the agent is idle. There is no hosted API, no paid tier, and no managed runtime — you own the infrastructure entirely, which means you also own the uptime.

    FreeOpen SourceSelf-hostedVerified Jul 19, 2026
34. 

    34. Tessera

    Tessera operates as a deterministic evidence layer that sits between your agent and its outputs. Every claim the agent surfaces is linked to a specific source record; claims without that linkage are refused outright, not softened or hedged. Before any action executes, the agent drafts it from verified claims only and surfaces it for your review. The architecture is open-source under MIT and built to integrate with MCP-based agent setups. Where it breaks: teams that need the agent to synthesize across sources where no single record covers the answer will hit refusals that require data-model work to resolve.

    FreeOpen SourceAPISelf-hostedVerified Jul 8, 2026
35. 

    35. UFO

    UFO is an open-source fleet coordinator for local AI coding agents. You enroll machines as rovers, assign work through a hub, and each operation runs in an isolated worktree with its conversation history, telemetry, and artifacts attached — not scattered across tabs. The auto-detection layer reads which AI CLIs are installed on each rover and advertises their capabilities for dispatch, so you are not manually tracking which machine has Claude Code versus Codex. Public beta status means the rough edges are real: APIs shift, documentation trails the code, and production stability is a bet you are making early. Teams with tight reliability requirements will hit that ceiling before teams prototyping fleet patterns.

    FreeOpen SourceSelf-hostedVerified Jul 15, 2026
36. 

    36. USB AI Agent

    The project ships 13 tools — DuckDuckGo deep search, OSINT via Holehe (121+ sites) and Maigret (600+ platforms), file read/write, Python and shell execution, and persistent memory that saves directly to the drive. Everything runs locally, leaving zero traces on the host machine, which matters on Tails or air-gapped hardware. The autonomous loop lets the agent decide which tools to call and in what order without you directing each step. Where it breaks: the repo has 2 stars and 4 commits, which signals a solo early-stage project with no documented community, no issue history, and no validation of the tool-calling loop at scale or across edge-case inputs.

    FreeOpen SourceSelf-hostedVerified Jul 24, 2026
37. 

    37. Katra

    Katra is self-hosted memory infrastructure: drop it on any Docker-capable machine, point your MCP-compatible agent at it, and you get episodic recall, semantic search, knowledge graphs, and temporal analysis without rebuilding your agent. The architecture is a single deployable unit — the vendor describes it as a 'memory appliance' — which means setup friction is low for teams that already run Docker or Helm on AWS. Where it breaks: Katra is memory infrastructure, not an agent runner, so teams expecting built-in task planning or tool execution will need to wire those themselves. The project is early-stage with five stars on GitHub and no reported production deployments in public community channels, which means you are taking on the role of early adopter rather than stepping into a proven stack.

    FreeOpen SourceAPISelf-hostedVerified Jul 1, 2026
38. 

    38. BrAIn

    Built on NATS as its messaging backbone, brAIn distributes agent nodes across hardware and wires them together through a reactive event bus — so an agent fires when something happens, not when a scheduler decides it should. Each node can carry its own UI, which means you monitor individual agents in context rather than reading logs and guessing. The architecture is documented (ARCHITECTURE.md, AGENTS.md), MIT-licensed, and ships with Docker and a monorepo package structure, so self-hosting is the intended path. The project is early-stage with 3 stars and 282 commits from a solo maintainer, which means production hardening and community support are things you contribute rather than consume.

    FreeOpen SourceSelf-hostedVerified Jul 9, 2026
39. 

    39. HART OS

    HART OS is an open-source, Apache-2.0 multi-agent runtime built on AutoGen that runs autonomous agents across a crowdsourced compute network, routes tasks through gossip-based federation, and keeps humans in the approval chain by design. The Recipe Pattern is the sharpest production differentiator: agents learn a task once in CREATE mode, then replay it in REUSE mode without repeating LLM calls — the vendor states up to 90% faster execution on trained tasks. Budget gating and compute escrow prevent any single node from absorbing costs for others. Where this breaks down is in ecosystem maturity: no comparable alternatives are listed in the market, documentation is structured but thin in places, and teams building beyond the Nunba bundled distribution will be navigating architecture that is still finding its production footing.

    FreeOpen SourceAPISelf-hostedVerified Jul 28, 2026
40. 

    40. tutti

    The core idea: instead of agents exchanging summaries, they share a live project state. Codex sees exactly what Claude changed, what's running, and what's pending — no copy-paste required. The `@` reference system lets any agent or teammate pull from any file or conversation in the workspace without re-uploading. A GUI control center surfaces every pending approval and running task in one view, so you sign off without opening a terminal. The ceiling appears when your workflow involves agents outside Tutti's supported roster or when you need fine-grained infra control — the platform is built for GUI-driven coordination, not headless pipeline automation.

    FreeOpen SourceVerified Jul 14, 2026
41. 

    41. Agency Agents

    The project is a MIT-licensed, self-hostable collection of pre-defined agent definitions organized by domain — engineering, marketing, product, design, and more — built to be activated inside Claude Code, Cursor, and similar AI coding tools. Each agent carries a defined personality, a stated process, and expected deliverables, so the session opens with role context already loaded. The differentiator is breadth plus specificity: you are not configuring a blank agent; you are picking a specialist with an opinionated approach baked in. The ceiling appears when your workflow requires branching between agents at runtime or dynamic handoffs — the repo defines agents, it does not orchestrate them. Teams needing cross-agent coordination wire that logic themselves on top.

    FreeOpen SourceSelf-hostedVerified Jul 4, 2026
42. 

    42. Auto Learning Agents

    The platform runs anywhere Docker runs: clone, add keys, and agents come online with their own schedules, models, and instructions. Every turn is embedded and topic-classified; agents query that shared memory bank before acting, so behavior improves with use rather than resetting on each session. Approval queues guard any outward-facing work — you review before it ships — and a single toggle pauses all autonomous activity. The Elixir/OTP supervision tree handles restarts without manual intervention, which matters when you need agents running overnight without babysitting. The extension story is custom nodes and tools in code, so teams that need behavior the built-in config cannot express will be writing Elixir or wrapping external services.

    FreeOpen SourceSelf-hostedVerified Jul 4, 2026
43. 

    43. AutoGPT

    The AutoPilot interface lets you describe a workflow in plain language and get a running agent without touching a flow diagram — which means a non-technical PM can ship something before an engineer has opened the docs. When the logic gets complicated, the Builder canvas lets you drag, branch, and loop at the block level, with live input/output inspection at every step. The Dashboard surfaces spend per agent and flags anything waiting on your input, so nothing runs invisible. The Marketplace gives you community-built agents with real run counts, skipping the blank-page problem entirely. Where it strains: teams reporting that complex multi-step branching inside the visual canvas hits readability limits, and the self-hosted path requires managing your own infrastructure when the hosted platform's scale limits bite.

    PaidOpen SourceAPISelf-hostedVerified Jul 1, 2026
44. 

    44. Loma

    Loma sits across your tools — Slack, docs, CRM signals — running agents that handle pre-meeting briefs, RFP responses, bug triage, and onboarding health checks without waiting to be asked. The differentiating claim is the context layer: every resolved ticket, closed deal, and fixed bug is stored as a pattern or skill that future agents draw on, so day 100 is meaningfully faster than day 1. Self-hosted under Apache-2.0, it supports Claude, GPT, and Gemini with swap-anytime routing. The vendor states agents complete RFP questionnaires at ~95% coverage, flagging the remainder for human review. Where it strains is in the gaps the scraped content leaves open — enterprise auth, SLA guarantees, and mature operational tooling are not documented.

    FreeOpen SourceSelf-hostedVerified Jul 2, 2026
45. 

    45. Rate A Human

    The workflow is deliberately minimal. You point your AI agent at a plain-text file at rateahuman.xyz/llms.txt, ask it to leave a rating, and the resulting star score plus a short written review appears on a public leaderboard. Reviews include a numeric rating, a paragraph of prose from the model, and a set of trait tags like 'direct', 'demanding', or 'laconic'. There is no API, no self-hosting option, and no structured data export — what you see on the leaderboard is what you get. The site is a novelty product, not an evaluation infrastructure layer, and its utility ceiling arrives the moment you want to do anything programmatic with the output.

    FreeOpen SourceVerified Jul 21, 2026
46. 

    46. Reference MCP

    Reference is a local MCP server that reads session transcripts and memory files — CLAUDE.md, AGENTS.md, and similar — from multiple AI coding tools, then exposes that history to whichever agent is asking. Register it once in each tool's MCP config and any agent can query what any other agent did before. The architecture is passive: Reference does not plan, execute, or chain tasks — it answers lookups. That scope is the point, and also the ceiling. Teams with more than a handful of tools, or who need structured, queryable memory rather than transcript search, will find the file-scanning approach starts to feel fragile as session volume grows.

    FreeOpen SourceSelf-hostedVerified Jul 1, 2026
47. 

    47. Skill Federation

    Skill Federation runs locally on your machine and connects to a catalog of over 100,000 vetted skills. When an agent hits a gap, it surfaces matches in milliseconds — each one license-checked, security-scanned, and provenance-tracked — then waits for your approval before installing into .claude/skills/. The benchmark evidence from the vendor is specific: a bare Claude Code agent solves 17.5% of SkillsBench tasks; with Skill Federation retrieving the top match, that climbs to 22.8%, roughly closing 27% of the gap to a hand-crafted ideal skill. The privacy boundary is narrow by design — only an abstract wish crosses the wire, never your code, plan, or outputs. The hard ceiling is integration breadth: Claude Code is supported, with Codex, Cursor, and Gemini listed as coming.

    FreeOpen SourceSelf-hostedVerified Jul 3, 2026
48. 

    48. SOCBench

    The published detection module runs three frontier LLMs across four analyst personas against 1,205 labeled NetFlow units, scoring each provider-persona pair on F1 (per-flow, per-pair, per-host), verdict accuracy, cost per alert, latency, and completion rate. That scoreboard lets you stop trusting marketing and start comparing models on telemetry that resembles what a real monitoring tier sees. The ceiling is visible immediately: detection is live, but triage, investigation, threat hunting, detection engineering, and threat intelligence are all roadmap items. If your evaluation need extends beyond binary flow classification, SOCBench does not cover it yet — and the roadmap carries no committed dates.

    FreeOpen SourceVerified Jul 8, 2026
49. 

    49. Two-tier-memory

    The library implements what the repo calls the 'two-tier fix': structured storage in a local SQLite database, with semantic or keyword queries pulling back only the relevant rows instead of the entire memory corpus. The core workflow is a single Python file and a SQL schema — add a memory, query a memory, done. It runs entirely on-device with no external API calls. The wall you hit is expressiveness: the schema is fixed, so teams with complex memory taxonomies end up forking the schema or layering their own abstraction on top. At that point you are maintaining a fork.

    FreeOpen SourceSelf-hostedVerified Jul 6, 2026
50. 

    50. GOAT 2.0

    GOAT2 runs a Telegram-facing multi-agent system on top of async DAG execution, with a three-tier memory stack — Redis for fast session state, ChromaDB for vector retrieval, and Letta for longer-horizon behavioral learning. The DAG runner means agents can execute in parallel where dependencies allow, rather than waiting in a serial queue. The modular layout — separate directories for agents, orchestrator, memory, plugins, registry, and tools — means you can swap a backend without rewriting everything else. The wall appears when you need a non-Telegram interface: the docs describe Telegram as the primary entry point, and rerouting to another frontend requires you to rebuild the interface layer yourself. Teams that need a REST API or web UI will be adding code before they ship anything.

    FreeOpen SourceSelf-hostedVerified Jul 4, 2026
51. 

    51. Jargo

    Jargo handles the full audio path: WebRTC in, a streaming transcription-to-reasoning-to-speech pipeline with turn-taking and barge-in, then audio back out — conforming to the RTVI protocol so existing clients drop in without rewrites. Go's goroutine model means hundreds of concurrent audio sessions don't share a global lock, which is the architectural argument for the whole project. The catch is printed in the README itself: this is early-stage, APIs are unstable, and betting a production system on it before the interfaces settle is a real risk. Teams that need a stable, documented voice pipeline today will find more mileage in Python-based alternatives while this matures.

    FreeOpen SourceSelf-hostedVerified Jun 28, 2026
52. 

    52. Ornith-1.0

    The model family spans 9B-Dense through 397B-MoE, all post-trained on Gemma 4 and Qwen 3.5 with a reinforcement learning framework that jointly optimizes the agent scaffold and the solution rollouts it produces. The vendor states benchmark performance on Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw at the top of the open-source range for each size tier. There is no hosted API — you pull the weights and run inference yourself, which means your infra team owns the deployment stack from day one. The 397B-MoE variant requires hardware that most teams do not have on-hand, so realistic entry for self-hosted production starts at the 31B-Dense tier. Community activity is early-stage: ten commits in the repository, six open issues, no closed pull requests.

    FreeOpen SourceSelf-hostedVerified Jun 30, 2026
53. 

    53. Agent Governance Toolkit

    Policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous AI agents.

    FreeOpen SourceAPISelf-hostedVerified May 1, 2026
54. 

    54. Agent Island

    Built by the Stanford Digital Economy Lab and described in arXiv paper 2605.04312, Agent Island puts language models into a shared environment and measures strategic behavior — not just task completion. The benchmark exposes gaps that standard evals miss: can a model read the room, shift alliances, and avoid being outmaneuvered by another agent? The interface exposes play and log views so researchers can inspect run-by-run behavior. Where it breaks: there is no API, no self-hosted option, and no published code repository, so teams cannot integrate Agent Island into a CI pipeline or adapt the environment to their own agent design.

    FreeOpen SourceVerified Jun 20, 2026
55. 

    55. AgentArk

    The vendor describes AgentArk as a 'secure-first, self-learning' agent runtime written in Rust, deployable via Docker on your own infrastructure. Core capabilities include scheduled automations, conditional watchers, trace logging with drift detection, and guard layers that require your sign-off before an agent acts. Context compaction — distilling what agents have seen from browser data and tool outputs — is built in, which matters when long-running agents start blowing past context limits. The self-evolution component (GEPA) is documented in the roadmap and architecture files, but community adoption is early — the repository shows six stars and zero forks at the time of curation. Teams that need a production-grade, battle-tested multi-agent backbone with existing integrations will hit the ecosystem ceiling fast.

    FreeOpen SourceSelf-hostedVerified Jun 22, 2026
56. 

    56. agentmemory

    Orbit is an open-source agent orchestration harness that wraps coding agent runs in bounded, dependency-ordered tasks, then gates task completion on real validation: tests, lint, and type checks must pass before an orbit closes. Every run produces structured JSON artifacts — agent output, rubric scores, accept/iterate/stop recommendations, and a human-readable progress log — so you have a trail to review, not just a diff to guess at. It runs against Claude, Codex, Cursor, or any agent that speaks JSON over CLI. The demo runs without an API key, which matters when you're evaluating whether it even fits your workflow. Where it strains: teams who need a web UI, multi-agent parallelism, or cloud-managed infrastructure will hit the limits of an intentionally small CLI harness fast.

    FreeOpen SourceSelf-hostedVerified Jun 7, 2026
57. 

    57. Agnt

    AGNT is a local-first agent operating system built around an AGI loop: the agent executes a step, evaluates the result, and re-plans before moving forward — without you steering each decision. Persistent memory and skill layers mean context survives across sessions, not just within a single run. The visual workflow designer handles repeatable paths; goal-mode hands the agent an objective and lets it figure out the steps. Self-hosted deployment with Docker keeps data on your own infrastructure, which matters when your legal team has opinions about where prompts and outputs live. The custom license — not OSI-standard — is the detail that stops procurement at some organizations before the first demo.

    PaidOpen Source$0 or $333/year per additional user for hosted versionAPISelf-hostedVerified Jun 9, 2026
58. 

    58. Alma

    Alma stores facts and preferences — name, role, working style, answer preferences, current context, principles — as a self-model any MCP-compatible agent can read at session start. The data stays on your machine; no hosted account, no vendor lock-in. Access is scoped, so an agent can read the slice it needs without touching the full store. Every durable write goes through an event log, which means changes are auditable and can be reversed. The project is explicitly labeled experimental by the maintainer, so APIs are unstable and behavior can change between commits.

    FreeOpen SourceAPISelf-hostedVerified Jun 24, 2026
59. 

    59. Atizar

    Atizar is an open-source, TypeScript-native framework for building agent workflows where humans stay in the loop before consequential actions execute. The core pattern: agents plan and gather, then pause for a sign-off before anything ships — emails send, records update, data moves. That approval gate is architectural, not bolted on after the fact. The self-hosted option means client deliveries stay off third-party infrastructure. Where it gets tight is documentation depth — the README carries most of the guidance, which means teams building complex branching logic are reading source code before long.

    FreeOpen SourceAPISelf-hostedVerified Jun 22, 2026
60. 

    60. AutoGPU

    The repo describes autonomous agents writing RTL, running it through real EDA tools, reading timing and layout reports, and revising the design — iterating without a human in the seat for each pass. The documented target is small systolic array architectures, specifically matrix-multiply accelerators; the codebase includes ISA definitions, physical design configs, and golden reference models. At that constrained scope, researchers report the agent loop closes. Scale the design complexity beyond what the existing module hierarchy covers and the agents lose the plot — the feedback loops that work for a mac array do not generalize to a multi-block SoC. Teams pushing past the documented scope end up writing their own agent scaffolding on top, at which point AutoGPU is a reference rather than a runtime.

    FreeOpen SourceSelf-hostedVerified Jun 11, 2026
61. 

    61. AutoLang

    Orbit wraps each agent run in a bounded loop: it pulls one task from a dependency-ordered backlog, hands it to whatever agent you've wired up, runs tests, lint, and type checks, and refuses to close the task until validation passes. Every run produces structured JSON — what the agent returned, how it scored against a rubric, whether a human should accept or re-queue. That audit trail is the point. The ceiling appears when your workflow needs anything beyond task-level sequencing: parallel agent execution, real-time dashboards, or integration with existing CI pipelines requires you to build the glue yourself.

    FreeOpen SourceSelf-hostedVerified Jun 2, 2026
62. 

    62. Autonomy

    The core loop — AgentLoop — runs up to a configured step ceiling, selects from 15 bundled procedural skills, ranks candidate actions across five weighted dimensions using beam search, executes through ActionGateway with LOW/MEDIUM/HIGH risk labels, then evaluates and learns. Every event in that chain is stored via event sourcing, so the full run is replayable. The learning loop drafts new skills after a successful run and queues them for review rather than auto-applying them. The wall appears when you need agents running in parallel or sharing state across concurrent sessions — the architecture is single-loop, single-goal. Teams that outgrow that model start wiring external orchestration around it.

    PaidOpen SourceFree Trial · 7 days$75/moAPISelf-hostedVerified Jun 22, 2026
63. 

    63. BGE-M3

    BGE is a family of open-source embedding and reranking models from BAAI, released under MIT license with weights available on Hugging Face and PyPI, designed to run entirely on your own infrastructure. The core workflow is straightforward: generate dense embeddings, index them in a vector database, and optionally layer in sparse or multi-vector retrieval for hybrid search. Multi-lingual retrieval is a documented strength, with cross-lingual matching working across language pairs without requiring parallel training data. The ceiling appears when your domain is highly specialized — out-of-the-box embeddings on narrow technical corpora produce ranking quality that requires fine-tuning to fix, and that fine-tuning work lands entirely on your team.

    FreeOpen SourceAPISelf-hostedVerified Jun 9, 2026
64. 

    64. Browser Use

    Browser Use is an open-source Python library for autonomous web task automation using LLMs and computer vision. Teams use it to extract competitive data, fill forms at scale, and monitor page changes across hundreds of sites. The tool hits 89.1% success on standard benchmarks and comes with stealth browser support, CAPTCHA solving, and residential proxies across 195+ countries. The vendor also runs a cloud infrastructure option alongside the self-hosted library. Most production teams pair it with managed browser infrastructure and human approval gates for financial or sensitive actions. The sharp edge: LLMs can't reliably distinguish user instructions from webpage content, leaving agents vulnerable to indirect prompt injection attacks that succeed 24% of the time without defenses.

    PaidOpen Source$29/moAPISelf-hostedVerified May 8, 2026
65. 

    65. Build A Harness

    The tool is a visual canvas for designing agent harnesses: multi-layer structures that govern what an agent believes, what it can do, and how it recovers when something goes wrong. You draw a graph on the canvas, and FlowSpec compiles it to LangGraph, CrewAI, Mastra, or Microsoft Agent Framework — no rewrite when you switch. The node library covers 27 types across 14 execution and 13 harness categories, from a minimal three-node loop to an 11-layer stack with world model, five-tier control resolution, nine-layer verification, and six recovery strategies. Langfuse tracing is built into the architecture, so every decision is observable. The ceiling appears when your team needs a hosted managed runtime — the vendor states no hosted API exists, so you own the infrastructure from day one.

    FreeOpen SourceAPISelf-hostedVerified Jun 25, 2026
66. 

    66. Ciris

    CIRIS runs a signed reasoning agent on your phone or a home device, with no warehouse in the middle for the closest privacy circles. The vendor describes two paths: fully on-device using a small model like Gemma 4, or free hosted inference for phones that can't run a local model — both paths produce cryptographically signed outputs. Every claim the agent makes carries an ed25519+post-quantum signature, so you can audit it, revoke trust, and re-open any conclusion built on a bad source. The architecture depends on a 'social circle' data model; data in your innermost circles never sends the network message that would let anyone request it. Teams needing broad third-party integrations or a hosted API endpoint will find neither here.

    FreeOpen SourceSelf-hostedVerified Jun 12, 2026
67. 

    67. Command R7B

    Command R7B is a smaller language model optimized for tasks that don't require reasoning at the frontier—summarization, classification, instruction-following, and document analysis. Cohere positions it as the pragmatic choice for teams tired of paying for (or waiting on) 70B+ parameter models when a tighter, faster alternative works. It's free and open source, which means no API charges and full control over deployment. The real limitation: it will struggle on abstract reasoning, mathematical proof, or multi-step logic puzzles where 70B models shine. For enterprises choosing between this and proprietary APIs, the tradeoff is real but worth calculating.

    PaidOpen SourcePay-as-you-goAPISelf-hosted128k tokens context$0.04/1M inputVerified Apr 9, 2026
68. 

    68. Conversations in AI Coding Agent

    Orbit is an MIT-licensed, self-hosted harness that wraps a coding agent run in a bounded loop: it selects a task from a dependency-ordered backlog, hands off to whatever agent you plug in, runs tests and lint as a hard gate, and writes structured JSON artifacts that record exactly what happened. Every closed orbit leaves four files — agent output, rubric scoring, an accept-or-iterate recommendation, and a human-readable progress log. The demo runs without an API key, which means you can verify the mechanics before committing any credentials. The harness is agent-neutral by design; the vendor page cites Claude, Codex, and Cursor as examples. Where it shows its seams: Orbit is intentionally small, so teams needing a hosted dashboard, team-level access controls, or CI/CD pipeline integration will be writing that glue themselves.

    FreeOpen SourceSelf-hostedVerified Jun 4, 2026
69. 

    69. CopilotKit

    The core model is a React and Angular SDK that connects your existing frontend to whatever agent backend you're already running — LangChain, CrewAI, or a custom setup — via the AG-UI protocol, a bi-directional event stream the vendor describes as 'the general-purpose connection between a user-facing application and any agentic backend.' Agents render rich UI cards, forms, and widgets inline as they work, not just text responses. Thread and state persistence is handled automatically across sessions. The friction point arrives when your deployment target isn't a web surface: Slack and Teams connections are flagged as early access, which means you're betting on a roadmap, not a shipping feature. Teams with strict approval gates before agent actions can wire those checkpoints in, but the docs describe this as a configuration responsibility rather than a built-in guardrail system.

    PaidOpen Source$39/developer/monthAPISelf-hostedVerified Jun 9, 2026
70. 

    70. CoreAI Model Zoo

    The repo ships Qwen3.5, Qwen3.6, Gemma 4, GLM-4, and LFM variants already converted, verified against iPhone 17 Pro GPU and ANE, and downloadable from Hugging Face. Conversion code, known gotchas, custom Metal kernels, and a Swift runner are included so teams can replicate or extend the work rather than reverse-engineer it. The larger dense and MoE models — Qwen3.6-27B, Qwen3.6-35B-A3B, GLM-4.7-Flash — are flagged Mac-only, so iPhone deployment is constrained to the smaller quantized variants. There is no API, no inference server, and no tooling outside the Apple ecosystem; teams targeting Android, Windows, or server-side inference will find nothing applicable here.

    FreeOpen SourceSelf-hostedVerified Jun 13, 2026
71. 

    71. CrewAI

    CrewAI helps enterprises operate teams of AI agents that perform complex tasks autonomously, reliably and with full control. The open-source framework (free, self-hosted) defines agents with roles, goals, and backstories, orchestrating them through tasks; the paid AMP adds a visual Studio, deployment infrastructure, tracing, guardrails, and enterprise features. The framework was rebuilt from scratch to remove LangChain dependency; as of v1.14, it's fully standalone and works with any LLM provider. It's used by nearly half of the Fortune 500. But production friction is real: common Reddit advice is to start with CrewAI for speed and migrate to LangGraph when you hit scaling limits—reasonable for most projects. Users report that enthusiasm evaporates when running repeatedly on multiple components, and executing large SELECT queries overflows the LLM context window.

    PaidOpen SourceOpen-source free; CrewAI AMP paid tiers start at $99/monthAPISelf-hostedVerified May 6, 2026
72. 

    72. DBRX Instruct

    DBRX Instruct is a free, open-source large language model built by Databricks for instruction-following tasks in software development and enterprise applications. It uses a mixture-of-experts architecture to balance performance with efficiency, and integrates natively with Databricks' data platform—a meaningful advantage if you're already in that ecosystem. The model shows strong results on coding and reasoning benchmarks, but carries real limitations: no vision capabilities, a shorter context window than Claude or GPT-4, and less real-world adoption in mainstream enterprise settings. For teams deeply embedded in Databricks infrastructure, it's a compelling option; for everyone else, it remains a secondary choice.

    FreeOpen SourceAPISelf-hosted32.8k tokens context$1.2/1M inputVerified Apr 8, 2026
73. 

    73. DeepSeek V3

    A fast, chat-based, Mixture-of-Experts (MoE) model from DeepSeek.

    PaidOpen Source$0.14 per million input tokens and $0.28 per million output tokensAPISelf-hosted131.1k tokens context$0.28/1M inputVerified May 15, 2026
74. 

    74. Due Diligence Agents

    The tool runs parallel analysis across Legal, Finance, Commercial, Technology, Cybersecurity, HR, Tax, Regulatory, and ESG workstreams — domains that siloed consultants hand off sequentially, bleeding weeks in the process. Each agent cross-references findings against the others, so a revenue concentration risk in the commercial workstream gets flagged against the indemnification language in legal without a human manually connecting the dots. Outputs land in Excel and Word with citations intact, ready for an IC memo. The knowledge compounds across deal runs, so repeat buyers in the same sector start with context the first team had to build from scratch. The ceiling appears when your data room contains formats the parser does not handle cleanly — and at that point, teams are pre-processing documents manually before the agents ever see them.

    FreeOpen SourceSelf-hostedVerified Jun 1, 2026
75. 

    75. Eidentic

    The SDK centers on a temporal knowledge graph that tracks when facts were true, resolves contradictions, and consolidates between sessions — so the agent sharpens over time rather than accumulating noise. Durable runs, enforced cost ceilings, and CI-gated evals ship as part of the core, not as paid add-ons. The vendor benchmarks report 55.2% on LongMemEval versus 41.0% for full-context stuffing, and claims up to 39× fewer tokens per query. The gap shows up in support and long-running assistant workflows where session history compounds. At v0.1, the ecosystem is early — teams building anything outside the TypeScript path face a hard stop.

    FreeOpen SourceAPISelf-hostedVerified Jun 12, 2026
76. 

    76. Enforra

    Orbit is a harness that wraps AI coding agents — Claude, Codex, Cursor, any JSON-speaking CLI — in a bounded task loop: the agent runs, tests and lint decide whether the work passes, and every run leaves inspectable JSON artifacts whether it succeeds or fails. The evidence trail is the product. You get structured output describing what the agent returned, rubric scoring for task focus and diff signal, and a human-readable progress log. Where it breaks: Orbit does not plan, does not write tasks, and does not decide what to build next — it validates and records what other agents attempt. Teams that need autonomous end-to-end execution will hit that ceiling immediately.

    FreeOpen SourceSelf-hostedVerified Jun 1, 2026
77. 

    77. Enju

    Orbit structures agent work into discrete, dependency-ordered loops: one task per run, deterministic validation gates, and four output artifacts that record exactly what the agent returned, how the run scored against a rubric, and what should happen next. The demo runs without an API key, which means you can evaluate the harness itself before spending a single token. Where it gets constrained: Orbit is a harness, not a scheduler — it does not autonomously drive through a backlog or retry failed orbits on its own. Teams wiring it into CI pipelines write the outer loop themselves.

    FreeOpen SourceSelf-hostedVerified Jun 1, 2026
78. 

    78. GEDD

    The vendor describes GEDD as a release-readiness tool for AI product managers and domain experts. A PM loads realistic launch-risk scenarios, the domain expert reviews the agent in the shape of the actual task, names failure modes in their own vocabulary, and the session exits with a release report plus a validated evaluation set. That loop converts qualitative judgment into regression gates usable in CI/CD. The ceiling appears when you need programmatic API access — GEDD exposes none, so teams that want to pipe evaluation results into downstream automation build that bridge themselves. Setup requires local installation via pip and depends on sagemaker-mlflow, grounded-evals, and mlflow.

    FreeOpen SourceSelf-hostedVerified Jun 9, 2026
79. 

    79. Genomi

    The core workflow is four steps: install the agent harness, point it at your raw genome file on disk, build a local SQLite index, then ask questions through whichever AI agent you already run — Claude Code, Cursor, Gemini CLI, Goose, and others are listed as compatible. Pharmacogenomics, carrier status, polygenic risk scores, nutrigenomics, and ancestry PCA projection are all covered through distinct skill modules backed by ClinVar, PharmCAT, PGS Catalog, HPO, GenCC, and 1000 Genomes reference data. The privacy architecture is explicit: raw genome data stays on disk, and only the specific evidence snippets relevant to a query cross the boundary to whatever LLM handles the response. The vendor marks this as experimental and not for clinical use — which means researchers and privacy-conscious individuals exploring personal data are the intended audience, not clinical teams expecting diagnostic-grade output.

    FreeOpen SourceSelf-hostedVerified Jun 9, 2026
80. 

    80. Goose

    Goose runs as a desktop app, CLI, or embeddable API — built in Rust, so the performance profile is consistent across macOS, Linux, and Windows without a runtime you have to manage separately. The extension system connects to 70+ tools via the Model Context Protocol, meaning a workflow touching GitHub, Google Drive, and a database isn't stitched together with custom glue code — the standard handles the handoff. Recipes let you capture multi-step workflows as YAML configs and share them across a team or drop them into CI. Where the architecture shows its limits: complex conditional branching inside recipes is not the same as writing that logic in code, and teams building workflows that require dynamic decision trees at depth report dropping into Python extensions to compensate — at which point they are maintaining two systems. Community support is Discord-first; the vendor states no paid tier, so production SLA expectations need to be reset before an org-wide rollout.

    FreeOpen SourceAPISelf-hostedVerified Jun 20, 2026
81. 

    81. Gorchestra

    Gorchestra is a locally-run Go server that wraps multiple Codex and Claude coding sessions into a single interface, streaming every event in real time and persisting the full history to SQLite. You can inspect live file changes and git state mid-run, queue follow-up messages without interrupting the agent, and return to the exact same session state after a disconnect or restart. The tool is MIT-licensed, self-hosted, and carries no hosted backend. The ceiling appears when your workflow demands coordinating agents across machines or teams — everything here is designed around a single developer's local environment, not a shared multi-user setup.

    FreeOpen SourceSelf-hostedVerified Jun 22, 2026
82. 

    82. Halo

    HALO is an open-source Hierarchical Agent Loop Optimizer that ingests production execution traces and generates RLM (Reinforcement Learning from Mistakes) reports pointing at the specific harness code responsible for systemic failures. The core loop is: run your agents, collect traces, feed them to HALO, receive a structured critique, patch the harness. It installs as a desktop app via a one-line curl command or as a hosted option through inference.net. The tool is built around planning and execution trace analysis, so it rewards teams who already instrument their agents — if your traces are thin, the reports will be too. Teams with dense trace data get targeted code-level feedback; teams without it get generic signal.

    FreeOpen SourceSelf-hostedVerified Jun 25, 2026
83. 

    83. Hermes Agent

    The agent lives on your server — not a vendor's — and connects to Telegram, Discord, Slack, WhatsApp, Signal, and email simultaneously, so the same agent handles a Slack request in the morning and a scheduled backup at night. Persistent memory and auto-generated skills mean it accumulates institutional knowledge over time rather than starting cold on each invocation. Real sandboxing across Docker, SSH, Singularity, Modal, and local backends means you can isolate risky tasks without routing them through a third party. The ceiling appears when you need managed reliability guarantees: at v0.16.0 this is early-stage software, and self-hosted operations teams carry full responsibility for uptime, credential management, and model API costs. Teams that need SLA-backed infrastructure typically wire Hermes into a managed hosting layer — which adds operational overhead the framework itself does not absorb.

    PaidOpen SourceAPISelf-hostedVerified Jun 9, 2026
84. 

    84. Hermes Desktop

    Hermes Studio is an open-source, self-hosted dashboard that wraps Hermes Agent in a control plane: task scheduling, multi-agent coordination, memory and skill management, cost tracking, and an approval gate for actions you don't want running unsupervised. The vendor describes it as MIT-licensed with no paid tiers, which means every feature ships without a paywall. The architecture assumes you are already running Hermes Agent locally — Hermes Studio is the interface, not the runtime. Teams that need cloud-hosted infrastructure or agents that run without a local Hermes Agent install will hit that wall immediately.

    FreeOpen SourceAPISelf-hostedVerified Jun 9, 2026
85. 

    85. hermes-desktop

    The application handles installation and configuration of Hermes Agent through a desktop interface, which means you're not hand-editing YAML to connect tools or restart sessions. From the same window you manage skills, memory, messaging gateways, and scheduled tasks — the pieces that usually require separate configuration surfaces or CLI literacy. The 278 open issues on the repository signal an active but unpolished project, so expect rough edges on non-standard setups. Teams that need production-grade reliability or multi-user access will hit walls the desktop model was never designed for. This fits a solo developer or small team running a local agent workflow, not an ops team deploying to shared infrastructure.

    FreeOpen SourceSelf-hostedVerified Jun 19, 2026
86. 

    86. HermesBench

    OpenResume is a browser-based resume builder and parser that keeps all data local: nothing is sent to a server, no account is required. You fill in a form, the tool renders an ATS-optimized PDF in real time, and you download it. The parser side lets you drop in an existing resume and see exactly how an automated screener will read it — which fields it finds, which it misses. The tool handles one job well. It does not support multiple resume versions with branching tailoring logic, and teams needing bulk generation or API-driven output will find no hooks to connect to.

    FreeOpen SourceSelf-hostedVerified Jun 9, 2026
87. 

    87. Hezo

    Hezo runs a hierarchy of agents — CEO, Coach, Captain, workers — each isolated in its own Docker container, with your secrets never passed directly into agent context. Instead, an egress proxy swaps placeholders for real credentials only when the destination host matches an allowed list, and every substitution lands in an append-only audit log. The Coach agent reviews completed work and writes learned rules back onto workers, so repeated mistakes get corrected without you editing prompts by hand. The ceiling appears when you need agents to hit destinations outside the allowed-host list, or when your workflow requires branching logic the org-chart model doesn't express — at that point you're editing configuration that the docs describe but don't walk you through in depth.

    FreeOpen SourceAPISelf-hostedVerified Jun 25, 2026
88. 

    88. Hugging Face Spaces

    Orbit acts as a harness around any JSON-speaking coding agent — Claude, Codex, Cursor, or others — running one task per cycle, executing tests and lint checks to decide whether the work advances, and writing structured JSON artifacts for every run. The dependency-aware backlog keeps each task bounded so agents do not drift across scope. Where it breaks: Orbit is intentionally minimal, so teams expecting a hosted dashboard, a GUI, or built-in agent adapters beyond CLI-level integration will build those layers themselves. The artifact trail is machine-readable JSON and a markdown log — useful for audits, not for a non-technical stakeholder who needs a summary.

    FreeOpen SourceSelf-hostedVerified Jun 7, 2026
89. 

    89. Kikubot

    Each Kikubot container polls one IMAP mailbox, feeds incoming email into an LLM agentic loop with a configured tool set, and replies over SMTP. Multi-agent workflows emerge naturally: a coordinator agent emails specialists, specialists reply, threads become the audit trail. The architecture requires a running mail server, which adds operational surface area before a single agent does anything useful. Teams with no existing mail infrastructure will spend more time on SMTP/IMAP setup than on agent logic. When the email-as-bus metaphor stops fitting — high-frequency tasks, sub-second latency requirements, or webhooks that can't wait for a polling interval — this architecture forces a full redesign.

    FreeOpen SourceSelf-hostedVerified Jun 12, 2026
90. 

    90. Langflow

    Open-source visual builder for constructing AI agents and RAG applications via drag-and-drop interface with Python extensibility.

    PaidOpen SourceAPISelf-hostedVerified May 15, 2026
91. 

    91. Llama 3

    Llama 3 is a large language model family designed to handle standard NLP workloads—text generation, translation, summarization, and sentiment analysis—across a range of scales. Meta released it as open source, meaning you can download weights, fine-tune locally, or run it on your own infrastructure instead of hitting an API. The catch: while free to use, the model is young relative to Llama 2, and local deployment requires real hardware or cloud credits. For teams building production systems, this trades managed convenience for control and lower long-term marginal costs.

    FreeOpen SourceAPISelf-hosted8.2k tokens context$2.65/1M input
92. 

    92. Llama 3.2 90B Vision Instruct

    Meta's 90B multimodal large language model with vision capabilities, fine-tuned for instruction-following across text and image understanding tasks.

    Open Source128k tokens context$2.04/1M inputVerified May 7, 2026
93. 

    93. Llama 4 Scout

    Scout carries a 10M token context window, meaning you can feed it an entire codebase or a stack of legal documents in a single pass without chunking pipelines or retrieval hacks. Maverick trades raw context depth for stronger multimodal reasoning, handling interleaved image and text inputs through native early-fusion architecture rather than a bolted-on vision adapter. Both models ship as open weights, downloadable from Hugging Face after license acceptance, with no API bill required if you run them yourself. The ceiling appears at inference: the Mixture-of-Experts architecture demands hardware that most teams do not have sitting idle, and running Scout's full 10M context window in practice requires significant GPU memory that a standard cloud instance will not cover.

    FreeOpen SourceAPISelf-hosted131.1k tokens context$0.11/1M inputVerified Jun 4, 2026
94. 

    94. llayer

    The core idea is radical reduction: state lives in an append-only .jsonl history file, the context window is a jq stream reducer, and the agent loop is a while loop in bash. Because every component is a standard Unix text pipe, you can slice the history file to rewind agent memory and replay any point — a capability most agent frameworks make architecturally impossible. Debugging is grep and pv, not a proprietary trace viewer. The ceiling appears fast: complex tool chaining or parallel agent coordination does not emerge naturally from a bash pipeline, and teams building anything beyond a single-agent REPL will spend more time fighting shell quoting rules than building product.

    FreeOpen SourceSelf-hostedVerified Jun 24, 2026
95. 

    95. Lobu

    Lobu connects to over 50 data sources — HubSpot, Stripe, Zendesk, Snowflake, GitHub, and more — and builds a live memory layer that agents query on schedule rather than on demand. A 'watcher' definition tells the agent what to look for and when to pause for a human to sign off before anything ships. That approval-before-action model is what makes the autonomous scanning safe enough to actually run unsupervised. The ceiling shows up when your workflow needs logic that doesn't fit a watcher definition — at that point you're writing connector SDK code and maintaining it yourself. Teams with deeply custom data pipelines will feel that constraint before teams running standard SaaS stacks.

    PaidOpen SourceAPISelf-hostedVerified Jun 20, 2026
96. 

    96. LocalFlow

    The core loop is deliberately small: Orbit selects one dependency-ordered task, hands it to whichever coding agent you wire in, runs tests, lint, and type checks, and only closes the task if the agent can prove the work passed. Every run produces four artifact files — structured result JSON, rubric-scored evaluation, a review recommendation, and a human-readable progress log. That paper trail is what lets you compare two agents on the same task by diffing artifacts instead of re-running demos. The harness runs locally with no API key required for the replay demo, so there is nothing to provision before you can see it work. The ceiling appears fast on non-coding tasks — Orbit is built for code-output validation and nothing else.

    FreeOpen SourceSelf-hostedVerified Jun 3, 2026
97. 

    97. LoopTroop

    The tool orchestrates a local pipeline — LLM council planning, an iterative execution loop called Ralph, and OpenCode worktree isolation — designed for multi-file feature work where correctness matters more than turnaround time. Every ticket goes through an interview phase before a line is code is written, resolving ambiguities via adaptive question batches that the vendor describes as intentionally taking over an hour. You review diffs and sign off before anything reaches your main branch. The tradeoff is explicit: LoopTroop is slow by design. Teams treating it as a fast pair-programmer will be frustrated inside the first session.

    FreeOpen SourceSelf-hostedVerified Jun 24, 2026
98. 

    98. MagesticAI

    The platform runs a pipeline of specialized agents — Planner, Coder, QA — that hand off work through isolated Git worktrees, so each task gets its own branch and a bad run does not contaminate the main codebase. You monitor execution in real-time through a web UI, which means you are not staring at terminal logs hoping the right thing happened. The vendor describes cross-session knowledge retention, so the system carries context between separate task runs. The architecture supports multiple LLM providers, which means you are not locked to one API when costs shift. At 78 stars and 184 commits, this is early-stage software — community support is thin and the blast radius of an undocumented breaking change falls entirely on your team.

    FreeOpen SourceSelf-hostedVerified Jun 9, 2026
99. 

    99. Memharness

    The core premise is storing facts, not strings, with two independent time axes: when something became true in the world and when the agent learned it — so querying past agent states is a real query, not archaeology through logs. Everything lives in a single SQLite file, which means the storage layer makes zero LLM or network calls and stays auditable. Recall combines hybrid vector search and full-text search with a source-staleness signal, so older or superseded sources rank down automatically. Where it breaks: the SQLite backend is a hard ceiling for teams expecting distributed writes or high-concurrency production deployments. Teams hitting that ceiling will need to treat memharness as a pattern to port, not a service to scale horizontally.

    FreeOpen SourceAPISelf-hostedVerified Jun 19, 2026
100. 

     100. MemPalace

     Orbit wraps agent runs in bounded loops: it selects one dependency-ordered task, hands it to your agent, runs tests and lint and type checks, and only marks work complete if validation passes. Every run produces structured JSON artifacts and a human-readable progress log, so you are reviewing evidence instead of trusting output. The agent-neutral contract means you can swap Claude, Codex, or Cursor behind the same harness and compare structured artifacts across runs. The tool is intentionally small — it handles the validation harness, not the full development lifecycle. Teams with sparse test coverage will find the validation gates have nothing to enforce.

     FreeOpen SourceSelf-hostedVerified Jun 8, 2026
101. 

     101. Mind-expander

     The agent drives the canvas: it can run `npx mind-expander` in the background, load skill integrations, and build guided tours through architecture. You see the same graph the agent is reasoning about, which means review decisions and refactor plans are grounded in actual dependency structure — not the agent's approximation of it. That shared view is the differentiator. The ceiling arrives with language support: Rust and TypeScript are covered, the docs describe more language frontends as planned. Teams whose core services are in Go, Python, or Java will hit that wall on day one.

     FreeOpen SourceSelf-hostedVerified Jun 9, 2026
102. 

     102. Mistral

     Mistral offers a family of large language models ranging from the lightweight Mistral 7B to the more capable Mistral Large, accessible both as open-source downloads and via paid API. The company positions itself as the cost-conscious alternative to ChatGPT and Claude, with a free tier covering basic use cases but throttled requests that frustrate serious users. Pricing for the API starts around $0.14 per million input tokens—roughly one-third OpenAI's rate—making it genuinely cheap at scale. The catch: public API documentation remains sparse, and the free tier's limitations mean you'll likely hit a paywall faster than expected.

     FreeOpen SourceAPISelf-hosted32k tokens context$0.15/1M input
103. 

     103. Mistral Large 2

     Mistral Large 2 is a general-purpose language model trained to handle complex reasoning, code generation, and multilingual work at the scale enterprises need. It's free to use via API or self-host, sits in the same performance tier as proprietary models from OpenAI and Anthropic, and can ingest documents up to 128,000 tokens long. The core trade-off: it has a knowledge cutoff earlier than competitors and lacks serious vision capabilities, making it less suitable for tasks requiring current events or image understanding. For teams optimizing on cost and reasoning quality rather than breadth of modalities, it's a genuine alternative to paid tiers.

     FreeOpen SourceAPISelf-hosted262.1k tokens context$0.5/1M inputVerified Apr 8, 2026
104. 

     104. Mnemo

     Orbit wraps each agent run in a bounded loop: it selects a dependency-ordered task from your backlog, hands it to whichever coding agent you point at it, then runs tests, lint, and type checks before the task is allowed to close. Every run leaves structured JSON artifacts — what the agent returned, how the output scored against a rubric, and a human-readable recommendation to accept, iterate, or stop. The agent-neutral contract means you can swap Claude for Codex behind the same harness and compare artifacts instead of gut feelings. Where Orbit hits its ceiling: it is a harness, not a planner, so teams that need autonomous task decomposition or cross-repo coordination will be adding that layer themselves.

     FreeOpen SourceSelf-hostedVerified Jun 6, 2026
105. 

     105. Multi Agent Protocol for AI Scientist

     The protocol pairs a tool-using Scientist agent with a stateful advisor called Socrates that cannot execute code, cannot issue directives, and cannot answer questions — it can only ask them. The advisor must emit [APPROVED] before the Scientist proceeds to the next experiment, which means every plan gets interrogated before compute is spent on it. The vendor reports this lifted test scores on four of five MLE-bench Kaggle tasks, with an average gain of +55.9% over the Scientist running alone. The ceiling appears quickly outside benchmark-style research tasks: there is no API, no UI, and the protocol is designed around a specific two-agent structure that does not generalize to arbitrary pipelines without custom work.

     FreeOpen SourceSelf-hostedVerified Jun 25, 2026
106. 

     106. Myco Brain

     The core mechanic is deterministic writes: the application code writes facts to Myco's Postgres store, not the LLM, so every stored fact carries a source document, a confidence score, and a full audit trail queryable via brain_why. One MCP server exposes that memory to Claude Code, Cursor, Codex, Windsurf, and any other MCP-compatible client simultaneously — write from Claude Desktop, retrieve from Cursor, no sync step required. The vendor publishes a 500-question LongMemEval result and a recall@5 figure using a recency reranker, both on the full benchmark set. The hard ceiling appears when your agents need to act on what they remember — Myco stores and retrieves facts; it does not plan, route, or execute tasks, so orchestration logic lives elsewhere.

     FreeOpen SourceSelf-hostedVerified Jun 20, 2026
107. 

     107. OnBuzz

     OnBuzz, built by Loxia, lets you spin up multiple autonomous agents that collaborate on tasks, connect directly to LLM providers, and execute work without a cloud intermediary. The Apache-2.0 license and self-hosted design mean your data and your agents stay where you put them. It ships as an Electron app with installers and binaries, so setup does not require hand-rolling a container stack. The tool-use and inter-agent collaboration model is genuinely capable — agents can hand off tasks, run in parallel, and schedule work without you babysitting. Where it strains: the community repository has modest GitHub traction (33 stars at time of indexing), which means documentation gaps surface quickly and community debugging support is thin.

     FreeOpen SourceSelf-hostedVerified Jun 27, 2026
108. 

     108. OpenAgents

     OpenAgents positions itself as the coordination backbone for distributed AI agents. You get a hosted workspace (or self-host) where agents working on separate machines discover each other, share files and browser context, and coordinate via @mentions. Installation is one-liner: install the Launcher desktop app, point agents at a workspace token, and they join. The platform is open-source with an active but modest community. The technical surface is clean—agents register on the network, events flow between them, and context stays shared. The hard part surfaces later: when your agents are actually doing different things (some coding, some reviewing, some managing), orchestrating handoffs stays manual. This is SDK-first, not no-code. If you're building a research team of specialized agents or debugging scenarios where you need human eyes on agent reasoning in real time, the shared workspace genuinely reduces context switching. If you're running a single coding agent that sometimes needs to call another agent, you might be over-engineering it.

     FreeOpen SourceAPISelf-hostedVerified May 8, 2026
109. 

     109. OSymandias

     The project ships a self-hosted runtime built on FastAPI, Celery, PostgreSQL, Redis, RabbitMQ, and Qdrant, so you get job scheduling, DAG orchestration, shared memory, tool execution, and a real-time dashboard without stitching services together manually. A Python SDK lets you define agents, attach tools, and wire multi-agent plans through goal decomposition — the runtime handles the queuing and dependency resolution. That stack is genuinely useful for research pipelines or internal analysis workflows where you control the infra. The ceiling appears when you need a managed hosted option: there is none, which means your team owns every database migration, worker restart, and Redis failover.

     FreeOpen SourceAPISelf-hostedVerified Jun 19, 2026
110. 

     110. Patina

     Orbit wraps each agent task in a bounded loop: the agent works, validation runs (tests, lint, type checks), and the task only closes when the checks pass. Every loop leaves structured JSON artifacts — what the agent returned, how it scored against a rubric, and a human-readable recommendation to accept, retry, or stop. This makes agent runs auditable after the fact, not just observable in the moment. The ceiling appears when your project needs multi-agent coordination or a hosted execution layer — Orbit is deliberately narrow, self-hosted only, and ships no managed runtime.

     FreeOpen SourceSelf-hostedVerified Jun 8, 2026
111. 

     111. penguinAI

     The tool runs conversational AI character chats, free with no gating on features. A Finite State Machine tracks emotional arc across each session, so characters shift between sarcastic, nervous, dramatic, and curious rather than defaulting to a single tone. The vendor's own benchmarks rate it above GPT and Claude on emotional variety and character consistency — though those benchmarks use a mix of human raters and an LLM judge, so treat them as directional. There is no API, no self-hosting path, and no way to wire these characters into an external product. What you get is the chat surface, and nothing else.

     FreeOpen SourceVerified Jun 19, 2026
112. 

     112. Preseason.ai

     Orbit sits between your backlog and your coding agent, selecting one dependency-ordered task at a time, running the agent, then forcing the result through tests, lint, and type checks before marking the task done. Every run writes structured JSON artifacts — what the agent returned, how the output scored against a rubric, whether a human should accept or iterate — so you are reviewing evidence, not trusting a diff. The agent-neutral contract means you can run Claude, Codex, and Cursor against the same task and compare artifacts instead of impressions. The harness is intentionally minimal; it does not schedule, it does not host, and it does not manage secrets — which means the moment your workflow needs cross-repo coordination or cloud execution, you are writing the glue yourself.

     FreeOpen SourceSelf-hostedVerified Jun 8, 2026
113. 

     113. Proctor

     Proctor wraps each agent execution in a Linux sandbox that cuts off access to hidden tests, fix history, and network egress, so the agent cannot read the answers before producing them. After the run, it produces a cryptographically signed verdict bundle that a third party can verify without re-running anything. The signing and forbidden-access timeline together mean cheating leaves a detectable trace. The tool targets researchers and benchmark maintainers on Linux — it is not a hosted service, carries no API surface, and requires you to operate your own infrastructure. Teams with Windows-only CI pipelines or no Linux sandbox provisioning hit an immediate wall.

     FreeOpen SourceSelf-hostedVerified Jun 24, 2026
114. 

     114. ProData AI

     Orbit is an open-source harness that wraps AI coding agent runs in a fixed loop: pick a task from a dependency-ordered backlog, run the agent, validate the output against tests, lint, and type checks, then record structured evidence before the task closes. Nothing advances without proof. Each run produces four artifact files — agent output, rubric scores, a recommendation, and a human-readable log — so you can inspect exactly what happened without replaying the whole session. The harness is agent-neutral; Claude, Codex, Cursor, or any JSON-speaking CLI plugs in behind the same contract. The ceiling appears quickly on teams who need anything beyond the validation-gate model — custom orchestration, parallel agent execution, or UI-driven workflow design are not in scope.

     FreeOpen SourceSelf-hostedVerified Jun 6, 2026
115. 

     115. Qwen

     Qwen covers text generation, coding assistance, multimodal understanding, and reasoning tasks across a range of model sizes, all under Apache-2.0 licensing, which means you can run it locally, fine-tune it, and ship it in a product without negotiating an enterprise agreement. The architecture is a Transformer decoder, so the fine-tuning toolchain your team already knows applies directly. Multilingual capability is a documented design goal, not a side effect, making it a practical choice for teams building outside English-first markets. The Qwen Studio interface offers free access for experimentation, while production-scale API usage routes through Alibaba Cloud — meaning your infrastructure story depends on which cloud you already operate in. Teams needing sovereign deployment or cost-controlled inference can self-host, but that path requires operational capacity the vendor does not manage for you.

     PaidOpen SourceAPISelf-hostedVerified Jun 24, 2026
116. 

     116. Qwen2.5 72B

     Qwen2.5 72B is a free, fully open-source large language model built by Alibaba that you can run on your own hardware. It competes directly with Claude and GPT-4-class models on reasoning, code generation, and math—areas where most open alternatives historically lag—while supporting 128,000 token contexts and multiple languages. The catch is computational: you'll need serious GPU investment (roughly $200k+ in hardware) to run it at scale, and like all LLMs, it has a knowledge cutoff and may need customization for niche domains. For organizations that can afford the infrastructure, it eliminates per-API-call costs entirely.

     FreeOpen SourceSelf-hosted32.8k tokens context$0.12/1M inputVerified Apr 8, 2026
117. 

     117. RoBrain

     RoBrain sits between your team's AI coding tools — Claude Code, Cursor, Copilot, Codex CLI — and a shared Postgres instance, capturing not just decisions but the alternatives your team ruled out. An MCP server runs inside the editor and surfaces relevant history before the agent acts; a batch Synthesis scan reads the whole corpus on a schedule to flag contradictions and drift that no single session would catch. That cross-session contradiction detection is where it separates from alternatives that only check at insertion time or silently delete the losing decision. Self-hosted on Apache 2.0 with your own Postgres; cloud extraction and the Planning API are paid-only features.

     PaidOpen SourceAPISelf-hostedVerified Jun 1, 2026
118. 

     118. RunbookHermes

     The agent runs multi-signal diagnosis across observability data, builds a root-cause hypothesis, and generates or updates runbooks from what it learns — so the next incident with the same failure pattern starts from a documented baseline instead of a blank slate. The approval-gated remediation workflow means automated action doesn't ship without a reviewer, which matters when the blast radius is a production service. Where it breaks: the repo is five commits deep with zero open issues, which signals early-stage software, not battle-hardened infrastructure. Teams with complex multi-service topologies will hit integration gaps before the agent's reasoning does. Self-hosting is required, so operationalizing this adds a deployment and maintenance surface your platform team owns.

     FreeOpen SourceAPISelf-hostedVerified Jun 9, 2026
119. 

     119. Skawld

     The SDK runs on Node.js 18+ and Bun 1.1+ as an ESM-only package, so it fits cleanly into modern TypeScript projects without a build-step fight. The vendor describes a minimal setup as a single `Agent` instantiation with a provider, a tool set, and a session — you are running a streaming agent loop in under a dozen lines. Where it starts to strain is on the documentation side: the README is thin, full docs live off-repo at skawld.com/docs, and community reports are sparse given the early star count. Teams who need battle-tested enterprise support or a large ecosystem of pre-built integrations will hit that ceiling fast.

     FreeOpen SourceAPISelf-hostedVerified Jun 9, 2026
120. 

     120. Strands Shell

     The core pattern is tight: decorate a Python or TypeScript function with `@tool`, pass it to an `Agent`, attach hooks that fire before or after each tool call, and the agent runs its loop. The `BeforeToolCallEvent` hook lets you inspect the tool's name and input — and cancel the call with a message if your conditions aren't met. That's not a workaround; it's the documented pattern. Where the framework gets quiet is multi-agent coordination — the docs describe single-agent tool loops clearly, but teams building agents that hand off to other agents will find precious little guidance on failure recovery between hops. When that gap bites, teams layer their own orchestration logic on top, which means maintaining that logic themselves.

     FreeOpen SourceAPISelf-hostedVerified Jun 22, 2026
121. 

     121. SynapCores Agent

     The repo, published by SynapCores under MIT, routes all memory, retrieval, semantic tool selection, and generation through the SynapCores backend — one database as the entire brain. There is no LangChain, no separate vector store, no framework glue to audit or upgrade. The project ships a browser chat widget and a live debug sidebar so you can watch memory recall and tool routing decisions in real time. That transparency is the differentiating feature — and also the boundary: the agent's intelligence rides entirely on the SynapCores backend, whose self-hosted deployment requirements the repo does not fully document. Teams that need the backend running on-premise will hit that wall before they hit a code problem.

     FreeOpen SourceAPISelf-hostedVerified Jun 9, 2026
122. 

     122. Tab Council

     Orbit wraps agent coding work in a bounded loop: it selects a dependency-ordered task, hands it to whichever agent you've wired up, then requires passing tests, lint, and type checks before the task closes. Every run produces structured JSON — what the agent returned, how it scored against a rubric, and a human-readable progress log. Nothing advances on the agent's word alone. The ceiling appears when your workflow needs anything beyond single-task validation loops: multi-repo coordination, branching logic between tasks, or a hosted dashboard for non-engineering stakeholders all require you to build on top of Orbit yourself.

     FreeOpen SourceAPISelf-hostedVerified Jun 2, 2026
123. 

     123. Tabbit

     Orbit wraps agent execution in bounded, dependency-ordered tasks: one unit of work at a time, with tests, lint, and type checks acting as the gate before progress is recorded. Every run produces four structured artifacts — result JSON, rubric evaluation, a review recommendation, and a human-readable progress log — so code review has evidence instead of vibes. The agent-neutral contract means you can swap Claude, Codex, or Cursor behind the same harness and compare artifacts on identical task sets. The ceiling appears fast: Orbit is deliberately small, so teams that need scheduling across distributed workers or CI/CD pipeline integration will be adding that infrastructure themselves. It is a harness, not a platform.

     FreeOpen SourceVerified Jun 1, 2026
124. 

     124. Vmette

     The threat model vmette solves is concrete: prompt injection on a fetched web page, a malicious package in an AI-suggested install, or model output that does something you didn't intend — all of it lands inside the VM, not on your host. The isolation is hardware-level, not a container namespace that a determined process can escape. Because everything runs on-device, no agent output leaves your machine to a third-party cloud sandbox. The ceiling appears at the edges: vmette is macOS-only, and teams whose agents need to run on Linux servers or in CI pipelines will need a different isolation strategy.

     FreeOpen SourceSelf-hostedVerified Jun 9, 2026
125. 

     125. Z3r0

     Z3r0 is an open-source, self-hosted workbench where a coordinating agent (Z3r0/CSO) delegates to five specialist agents — code audit, recon, exploitation validation, reverse engineering, and cryptography — each scoped to a defined domain. Sessions run against a PostgreSQL-backed timeline log with replay, so long engagements survive interruptions and context window rollovers. WorkProject records tie every finding to authorized scope, targets, and sandbox bindings, which means the evidence chain stays intact when the model context doesn't. The wall appears when your engagement requires a specialist task not covered by the six fixed roles — there is no agent plugin system described in the docs, so teams extending scope are writing new agents from scratch.

     FreeOpen SourceAPISelf-hostedVerified Jun 11, 2026

Listings on this page are sourced and verified by the AIDiveForge data pipeline. AIDiveForge is editorially independent — no money changes hands for inclusion.