Start here
Understand MiniClaw, its current capabilities, and the shortest path to a local run.
MiniClaw is a small, complete, self-hosted personal AI agent with explicit boundaries. It puts Agent, Policy, Approval, Tools, Memory, and multiple messaging surfaces around one Python core so capability and control remain part of the same system.
Pick a path
- First run: use Install and run for dependencies, initialization, Doctor, and TUI startup.
- Understand execution: read Runtime to see how messages enter one AgentRuntime.
- Set permissions: read Security boundaries before choosing a Permission Mode.
- Connect a platform: read Surfaces before the platform-specific repository guides.
- Govern memory: read Memory for Markdown Truth and SQLite Projection.
Current, verifiable scope
MiniClaw currently exposes four surfaces—TUI, Feishu, Telegram, and Discord—with 18 built-in tools, 33 versioned Channel cases, and 15 versioned Automation cases. Automation is disabled by default.
Evidence boundary
Passing local fake-SDK scenarios and soak runs is an Implementation PASS. It is not automatically a real-platform Live PASS; consult current release evidence for each real-platform Live Gate.
Deeper engineering material
These docs stay focused on users. The complete scope, architecture, and evaluation evidence live on GitHub: