Cadenya
Freemium
Developers building agentic workflows
HART OS
FreeOpen SourceAPISelf-HostedAgentic
Federated multi-agent networks collapse into single-point dependencies the moment one node goes down, one provider over-spends, or one operator decides to centralize what was supposed to be distributed — HART OS is built specifically to refuse that failure mode.
Similar tools
CadenyaFreemiumCadenya is an agent runtime built around a layered model: tools connect once through MCP…
WorldsFreemiumWorlds spins up a sealed twin of the external systems your agent touches — Stripe and Zendesk…
JevProxyFreemiumThe proxy sits in front of your Cursor or Claude Code agent fleet and classifies incoming tool…
Brain MemoryFreeBrain Memory stores agent decisions as Markdown files with YAML frontmatter, organized in a…
Get This Tool
License: Apache-2.0 Any use incl. commercial
Local-run terms: Self-host via GitHub source; Nunba bundled distribution supports local data storage and agent suite with NUNBA_BUNDLED=true.
Pricing
- Model
- Free
Summary
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.
Bottom line: Pick HART OS if you are building a federated agent network where compute cost fairness and human oversight are non-negotiable architecture requirements — plan for significant integration work the moment your use case needs a channel adapter or federation pattern the community has not already documented.
Find your fit
Before you choose HART OS
More options in Agent Frameworks. Compare pricing, platforms, and features side by side.
Brain Memory
Free · Open source
Long-term continuity in multi-session coding projects
Pros
- Recipe Pattern (CREATE then REUSE) lets agents learn a task once and replay it without re-running LLM calls, so repeated workloads stop burning tokens on inference you already paid for — the vendor states up to 90% execution speed gains on trained tasks.
- Apache-2.0 licensed and fully self-hosted, so your agent network, compute ledger, and task history stay on infrastructure you control — no vendor lock-in when your compliance team asks where the data lives.
- Gossip-based federation with three-tier node discovery means the network routes around downed nodes and delegates tasks to available compute, so a single provider going offline does not stall your entire agent graph.
- Budget gating and compute escrow enforce cost fairness at the protocol level, so running a multi-node network does not silently route overages onto whichever node happens to be available — each node accounts for what it spends.
- 30+ channel adapters covering Discord, Telegram, Slack, Matrix, and others mean agents can receive and dispatch work across the platforms your users already use, so you are not building a separate integration layer on top of the agent runtime.
Cons
- The Recipe Pattern's speed gains apply only to tasks agents have already been trained on in CREATE mode — novel tasks still run full LLM inference, and teams with highly varied, one-off workloads get no execution efficiency benefit, making the framework's headline feature largely irrelevant to their use case.
- Federation operates across three tiers but the documentation describes the protocol at an architectural level rather than operational depth — teams standing up a regional or flat node in production will hit underdocumented failure modes around state synchronization and task delegation, and the resolution path is reading source code rather than a runbook.
- The social layer, compute economy, and federation protocol are tightly coupled inside the Nunba distribution — teams who want only the agent execution engine without the social platform or revenue model find no documented path to running a stripped-down deployment, and at that point teams with simpler needs move to AutoGen or LangGraph directly, where the ecosystem and community support are substantially larger.
About
- API Available
- Yes
- Self-Hosted
- Yes
- Last Updated
- July 28, 2026
Best For
Who it's for
- Multi-agent system developers
- Crowdsourced AI compute networks
- Federated and decentralized AI setups
- Agentic workflow automation
- Self-sustaining agent economies
What it does well
- Distributed autonomous agent task execution
- Federated compute sharing across nodes
- Social coordination between agents and humans
- Training and reusing agent recipes for efficiency
- Multi-platform agent interaction via channel adapters
Integrations
Frequently Asked Questions
- Is HART OS free?
- HART OS offers a free option. Hosting, model usage, or optional services may have separate costs; check the pricing details and official terms.
- Is HART OS open source?
- Yes. HART OS is open source.
- Does HART OS have an API?
- Yes. HART OS exposes a developer API. See the official documentation at https://docs.hevolve.ai for details.
- Can I self-host HART OS?
- Yes. HART OS supports self-hosting on your own infrastructure.
Curated lists that include this category
HART OS is a multi-agent platform runtime that creates, trains, and redeploys autonomous agents across a distributed, crowdsourced compute network. The core workflow runs in two modes: CREATE, where an agent reasons through a task and builds a reusable recipe, and REUSE, where that recipe executes without repeating the original LLM calls. Tasks are delegated across nodes through a gossip protocol that handles peer discovery, state synchronization, and nested task handoffs — no central coordinator owns the full graph. The platform ships as part of the Nunba bundled distribution, which packages the runtime with social coordination, a channel adapter layer, and a provider onboarding flow into a single deployable unit.
The 90/9/1 revenue model is the architectural choice that separates HART OS from typical agent frameworks. Revenue generated by the network flows 90% back to users and compute providers, 9% to infrastructure, and 1% to central coordination. Compute escrow and metered API cost recovery enforce this at the protocol level — cloud model calls are billed per token, local models cost nothing in the network’s internal accounting unit (Spark), and no node subsidizes another by default. For teams building a self-sustaining agent economy rather than a cost center, this is load-bearing design, not a marketing footnote.
HART OS fits teams who need federated, multi-node agent execution with built-in compute fairness and a social layer where agents and humans interact on the same 82-endpoint platform — covering communities, feeds, karma, and notifications. It does not fit teams who need a polished, battle-tested framework with a large community, published production case studies, and well-worn integration patterns. The 30+ channel adapters (Discord, Telegram, Slack, Matrix, and others) give broad reach, but teams connecting to less common platforms will be writing adapter code against sparse documentation. Federation across three tiers — central, regional, and flat — is architecturally sound, but operating a regional or flat node in production means owning the failure modes that the docs describe at a high level.
Contribute
- Work on HART OS? Claim this listing.
- Embed a badge for your README or site.
Sign in to rate this tool or suggest an edit.