Interface
A multi-agent AI system with self-correction, a ReAct loop and autonomous tool use. The agent plans tasks, calls tools, checks quality and retries if needed — with no user involvement.
Open on GitHub →The Worker doesn't just generate text — it autonomously calls tools, fetches data, saves results and forms the final answer.
Automatically discovers available models across 10+ providers in real time. The router picks the best one for the task.
The system can analyze its own code via read_source/write_source and make improvements. 15+ prompt versions archived.
Persistent vector memory with hot and warm tiers, a corrections archive and a user profile.
Files, browser, GitHub, shell, web search, notifications, system monitoring — the agent calls them autonomously.
Automatically discovers and connects to MCP servers to extend the toolset.
A task scheduler — the agent can run tasks on a schedule with no user involvement.
Successful corrections are distilled into reusable SKILL.md documents — failures of cheap models turn into a skill for future tasks.
Works as an OpenAI endpoint — connects to IDE plugins (Continue.dev and others). The Critic is two-tier: deterministic checks first, then LLM scoring.