AgentOS
Open-source TypeScript agent runtime whose agents can forge their own tools
Overview
AgentOS is an open-source TypeScript runtime for AI agents from Frame (Framers Lab, Inc.), published on npm as @framers/agentos (Apache 2.0 core, MIT extensions). Agents can forge JavaScript tools at runtime behind tests and an LLM judge, with persistent cognitive memory, six multi-agent orchestration strategies and 13 LLM providers.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 12 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with AgentOS
- Runtime tool forging in JavaScript, gated by test cases and an LLM judge
- Forged code run in node:vm or an isolated QuickJS WebAssembly executor
- Persistent cognitive memory with Ebbinghaus decay and eight named mechanisms
- Six multi-agent orchestration strategies
- Routing across 13 LLM providers, including local models, with fallback chains
- Guardrails at five security tiers, including PII redaction and prompt-injection detection
- Multimodal RAG and voice pipelines
AgentOS is unusual in two ways: agents can forge new JavaScript tools mid-task, gated by test cases and an LLM judge, and its memory layer models forgetting and recall rather than acting as a flat vector store. Its self-reported LongMemEval scores are strong, with a public harness but no independent rerun.
Development is fast and almost entirely one maintainer's work, and the default node:vm sandbox is not a security boundary. Treat it as promising for memory-heavy, long-running agents rather than a safe default.
What we found
AgentOS is an open-source TypeScript runtime for AI agents, published on npm as @framers/agentos. The core is Apache 2.0 and the agents, extensions and guardrails are MIT. It is built by Frame (Framers Lab, Inc.), which also runs the Wilds.ai game platform and the Paracosm simulation engine on top of it. The project is very active, with version 0.13.56 shipping on 11 October 2026 as the fifth release that day, and it is close to a one-person effort: the lead maintainer, Johnny Dunn, accounts for nearly 2,000 commits and the next human contributor for five. That matters more for long-term risk than any feature list.
Two ideas set it apart. When no tool covers a sub-task, an agent can write a JavaScript function with JSON Schemas for its input and output plus test cases; the tests run, a separate LLM judge reviews the result, and an approved tool joins that session's tool list. Code mode stays off until the host enables it, and the README says plainly that the default node:vm sandbox is not a security boundary, offering a QuickJS WebAssembly executor as the stronger option. The memory layer applies Ebbinghaus decay and eight mechanisms from cognitive science, among them retrieval-induced forgetting and reconsolidation, and an optional HEXACO personality model shapes how agents store and recall.
Around that sit six orchestration strategies, five guardrail tiers, multimodal RAG, voice pipelines and 13 LLM providers, including local models. The headline benchmark comes from the project itself: 85.6 per cent on LongMemEval-S with a router that sends most question types to gpt-5-mini, 84.2 per cent with gpt-4o alone, and 70.2 per cent on the harder M variant. The harness is public, but nobody independent has published a rerun, and earlier versions of this page credited the top score to gpt-4o.
There is no paid tier. Frame offers enterprise support and licensing by email, with no published prices.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
About this listing
This entry was compiled from publicly available data including AgentOS's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with AgentOS unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to AgentOS directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to AgentOS
More in Agents & Automation
