AI Agents Scoreboard (Sep 2026)

1

Famulor White-label Platform

The platform runs inbound and outbound phone calls, WhatsApp marketing sequences, booking flows, and multi-channel support from a single visual flow builder — no code required for the core path. The Milian AI Loop drives campaign retries, parallel dialing windows, and voicemail handling without manual intervention. Where it holds up: solopreneurs and agencies running repeatable call-and-message sequences. Where it hits a wall: teams that need self-hosted infrastructure for data residency beyond what EU-cloud hosting covers, or that need to fork the agent logic in ways the visual builder cannot express. At that point, teams either accept the hosted constraint or move to a self-hosted open-source stack.

Paid $0.14-$0.19 per minute after included credits; monthly plans from $29APIVerified Sep 16, 2026

11.4 score

2

AgentMesh.help

The platform is a REST-native task market where agents register in three API calls, browse open tasks, and submit deliveries without claiming or locking — multiple agents race the same task simultaneously. Budgets escrow at post time so agents know payment is real before they start, and settlement is instant when a poster picks a winner. The directory feature lets agents publish capabilities and receive direct invites, reserving a 48-hour exclusive window. The platform is early-stage — 12 registered agents and 30 completed tasks at the time of scraping — which means task volume is thin and an agent may burn cycles delivering work on tasks that already have a winner picked. Teams that need guaranteed task supply will find this unpredictable.

Paid APIVerified Sep 16, 2026

10.8 score

3

SendRoq

You describe your ideal buyer in plain English, and SendRoq's agents surface real LinkedIn profiles, generate personalized outreach, and run multi-step follow-up sequences until someone replies. The vendor states messages are tuned to your tone, and you can approve each one or let the agent run autonomously. For a solo founder trying to hit 50 conversations a month without a BDR, that loop closes fast. The ceiling appears when you need CRM sync, webhook triggers, or anything outside LinkedIn — the scraped page describes no API and no integrations with downstream tooling. Teams that need outreach connected to a broader revenue stack will hit that wall before the second quarter.

Paid $59/moVerified Sep 11, 2026

10.6 score

4

NoInfra

The vendor delivers pre-configured, hosted agents across a set of templated use cases: study summarization, job application tracking, spreadsheet cleanup, meeting prep, and sandbox code execution. You open a workspace and the agent is already running — no keys, no servers, no config files. The managed runtime abstracts token provisioning server-side, so users see a balance and status indicator rather than provider credentials. Where this model breaks: the compute ceiling is fixed per tier, and teams whose workloads outgrow the allocated vCPU and RAM have no self-hosted escape hatch — they either upgrade or leave.

Paid $19.99/mo and upVerified Jul 9, 2026

10.5 score

5

Cadenya

Cadenya is an agent runtime built around a layered model: tools connect once through MCP, OpenAPI specs, or existing endpoints; agents carry assigned tool sets, memory layers, and system prompts; objectives dispatch into a loop. The variation system is the differentiating structure — you run a canary agent against a different model or tool assignment alongside your default, compare scored feedback per objective, and promote or discard without touching infrastructure. Progressive tool discovery keeps full tool schemas out of the context window until the agent asks, which means fewer tokens consumed per request. The ceiling appears when your logic needs branching that the runtime does not expose as a configurable primitive — teams building conditional pipelines will reach that wall before teams running single-agent dispatch loops.

Paid $20 minimum monthlyAPIVerified Sep 14, 2026

10.2 score

6

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.

Paid Open source Self-hostedVerified Jul 21, 2026

9.9 score

7

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.

Paid Open source APIVerified Jul 23, 2026

9.7 score

8

EverMemOS

EverMemOS, built by EverMind, is a memory infrastructure layer that gives AI agents persistent, inspectable, and portable memory across sessions, platforms, and model providers. The vendor describes multimodal ingestion, so agents can encode not just text exchanges but structured context from multiple input types. Self-hosted deployments run under an Apache 2.0 license, which means teams with data-residency requirements can own the stack entirely. The ceiling appears when memory graphs grow dense — community reports suggest retrieval latency climbs before tuning is required, and teams building high-throughput customer support pipelines report needing to manage memory pruning manually. Teams that need memory to double as a full observability or analytics layer find they are adding a second tool alongside it.

Paid APISelf-hostedVerified Aug 16, 2026

9.7 score

9

Lunen.ai

A subject-matter expert describes what they want in plain language; Lunen drafts a structured execution plan with named tools, scoped data, and a schedule — no canvas, no YAML. Every MCP tool connection becomes a per-tool policy decision: allow it to run unattended, or pause for a human sign-off before each call. User actions and agent actions land in the same audit log, which means security reviews have a single trail to pull. The ceiling appears when teams need conditional branching between agent steps — the plain-language plan model does not surface that logic visibly, so complex multi-step dependencies require workarounds the interface does not directly support.

Paid Self-hostedVerified Jul 20, 2026

9.6 score

10

Nolo.Chat

The core idea: let your highest-capability model plan and coordinate, then hand execution to lower-cost agents running in parallel. The vendor describes patterns for coding delivery, parallel brainstorming, consensus deliberation, and video generation — all configurable by you, or auto-designed by the platform for a given task. Context persists across sessions, so agents carry rules and history forward instead of starting cold each time. Where this gets complicated is on the API side — Nolo offers no public API, which means you cannot slot it into an existing orchestration stack without going through its own client. Teams that need programmatic control over agent handoffs will hit that wall fast.

Paid Verified Sep 18, 2026

9.6 score

11

Reflexio

Reflexio sits between your agent and your logs, pulling corrections, failed paths, and successful outcomes into a learning loop that revises agent behavior over time. Each lesson is visible in a review interface where you can rewrite, approve, reject, or delete it — and a rejected learning stops being applied immediately. The self-tuning layer watches how each learning performs across real sessions and rewrites it from cases where it fell short, so behaviors sharpen with use rather than calcifying at capture. The ceiling appears when your team needs to audit the model generating those learnings at a granular level — Reflexio scores conversations against metrics you define, but the underlying methodology is the vendor's, not yours. Teams that need full control of the learning algorithm or on-premise data residency will hit that wall fast.

Paid APIVerified Sep 9, 2026

9.6 score

12

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.

Free Open source Self-hostedVerified Jul 13, 2026

9.6 score

13

AgentCaly

The core workflow is event-driven: you create or annotate a calendar event, and an agent runs a defined task at that time — pulling news, finding leads, checking flight fares, or drafting a blog outline — without you opening another app. The calendar-as-interface model removes the scheduling layer most automation tools require you to build separately. Where it strains is depth: agents suited to daily briefings and contact lookups hit their ceiling when tasks require conditional branching across more than a couple of steps. Teams that need complex multi-step decision logic will find themselves wanting a dedicated orchestration layer this tool does not provide.

Paid $15/month (PRO)Verified Jul 11, 2026

9.4 score

14

Agently

Agently connects to 100+ tools via OAuth and builds a live graph of your company's activity, then runs a set of specialized agents — Researcher, Revenue, Growth, Support, Ops, Briefer — coordinated by an orchestrator called Jarvis. Agents post Slack threads, recover failed Stripe charges, flag renewal risks, and ship formatted documents without waiting for a prompt. The output is artifacts — sheets, docs, decks, gated pages — not chat transcripts. The ceiling appears when you need conditional branching that goes beyond the predefined agent roles; the vendor describes no mechanism for custom agent logic or self-hosted deployment. Teams with non-standard workflows will feel the constraint.

Paid $69/moVerified Jul 15, 2026

9.4 score

15

bitroad

Bitroad is a marketplace where agents buy services from other agents — data labelling, GPU time, code review, research reports — inside spend caps you set before any money moves. The workflow is a single API key: your agent searches listings, commits funds to escrow, receives the deliverable, and releases payment only after you accept. The audit log captures every move. The ceiling is hard — requests over your per-job or per-day cap are blocked before a charge runs. Where it breaks: the seller side is the unknown variable. Disputed deliverables freeze escrow, but resolving that dispute is not an automated step.

Paid APIVerified Sep 14, 2026

9.4 score

16

Loop me in

The workflow is mechanical by design: an expert publishes a named loop with a fixed rate, the required context, and the specific judgment they return. When an agent hits that loop, it assembles context automatically and holds until the expert answers — from wherever they are. The model works cleanly for agents with clear pause points: brand reviews mid-code-generation, onboarding drop-off calls during product iteration, incident root-cause reads from logs. Where it strains is any run that needs real-time back-and-forth rather than a single bounded exchange, or any team whose agent framework does not support MCP hosts. The expert-as-contractor model is early — the vendor states V1.0.0, so directory depth and answer latency are open questions.

Paid Verified Jul 24, 2026

9.4 score

17

Scarlett.

The agent runs 24/7 across messaging channels, autonomously completing multi-step tasks: a patient asks about availability, Scarlett checks the schedule, books the slot, and fires a reminder — no staff action required. For a single-doctor practice or a business with two to ten staff, that coverage is the core value proposition. The vendor states HIPAA alignment, which matters the moment you are handling patient data in a regulated environment. The scraped page is thin on integration specifics, so verifying EHR or PMS connector depth before committing is non-negotiable. Multi-location teams should validate whether workflow logic scales cleanly across sites or requires per-location configuration.

Paid $97 USD / mo startingAPIVerified Jul 11, 2026

9.4 score

18

adris.tech

adris is a desktop app for Windows and Linux that bundles eight modules — AI agents, automation, a code editor, local model hosting, a credential vault, DNS-level threat blocking, cross-machine RAM pooling, and a shared knowledge graph — under one login. The agents (called Krew) research prospects and verify contacts in a live browser, then hand results directly to automations that push to Slack, Sheets, or Notion on schedule. Everything stores locally in SQLite; credentials never leave the device. The ceiling appears when you need a public API to connect adris to an existing internal system — the vendor does not list one. Teams that need to pipe agent output into a custom backend will hit that wall fast.

Paid from ₹0Self-hostedVerified Jul 20, 2026

9.2 score

19

Vora IQ - StartupOS for Founders

The core workflow starts with your idea or existing business: Vora IQ runs a viability analysis — market demand, competitor landscape, real signal — then generates a living roadmap that reprioritizes as your business moves. Agents covering strategy, finance, marketing, legal risk, and social content operate in the context of your specific business, so advice is not generic. The daily task queue is the most practical output: you open the tool and it tells you exactly what to do next. Where it shows strain is depth — a solo founder validating a SaaS idea gets real value, but a team that needs complex financial modeling with custom assumptions or nuanced legal review will find the agents hit their ceiling before a specialist would.

Paid $29.99/mo or $180/yrVerified Aug 16, 2026

9.2 score

20

Brain Memory

Brain Memory stores agent decisions as Markdown files with YAML frontmatter, organized in a directory tree you can browse in any file explorer — no opaque vector database, no embeddings you cannot audit. Strength decays on an Ebbinghaus exponential curve and rebuilds each time a memory is recalled, so the architecture you revisited three times stays sharp while the one-off experiment fades. The benchmark vendor cites shows 100% recall on a 1,000-distractor haystack where BM25 and vector retrievers both score zero — a meaningful gap for long-running coding projects. The tool is at v0.1.0, MIT-licensed, and ships as an npm global install. That version number is not a warning to ignore: the sleep consolidation pipeline and the cross-agent sync model are genuinely novel, which means the surface area for early-stage bugs is wider than a mature retrieval library.

Free Open source Self-hostedVerified Aug 14, 2026

9.1 score

21

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.

Paid Open source $0-$8/moAPISelf-hostedVerified Jul 22, 2026

9.0 score

22

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.

Free Open source APISelf-hostedVerified Jul 23, 2026

9.0 score

23

Nelieo NSP

NSP attaches to running JVM, V8, and CLR processes on Windows and exposes their internal object graphs at reported p50 latencies around 11ms — no rendering pipeline, no DOM traversal. The vendor's benchmark compares token consumption per task at roughly 580 tokens via NSP against ~14,850 tokens for a vision-model loop, which, if it holds at your workload, translates to a dramatic drop in per-workflow cost. The architecture is deterministic: agents invoke functions by memory address rather than locating UI elements, so a CSS drift or layout change does not break the execution path. The ceiling appears at runtime scope — NSP is Windows-only, covers V8, JVM, and CLR, and native-compiled binaries or runtimes outside that list get no structured access. Teams automating cross-platform or browser-agnostic workflows will hit that boundary quickly.

Paid APISelf-hostedVerified Aug 17, 2026

8.9 score

24

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.

Free Open source Self-hostedVerified Jul 23, 2026

8.8 score

25

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.

Free Open source Self-hostedVerified Jul 14, 2026

8.8 score