Experimental Research fork — APIs and behaviour may change without notice. Read disclaimer →

Picobot is a research fork of HKUDS/nanobot, focused on agent loop reliability and knowledge architecture.

Status

This project is experimental and under active development. Expect:

  • Breaking changes to APIs, config schema, and CLI flags without notice
  • Incomplete or evolving documentation
  • Features that work in some configurations but not others
  • No stability guarantees between versions

Intended audience

Developers who want to study or tinker with a small (~6k lines), readable AI agent loop. Not intended as a production dependency.

What's forked

Picobot diverges from nanobot in several areas:

  • ReAct loop — progress-gated continuation, stuck detection, sub-agent stop hook
  • Knowledge base — renamed WorkspaceIndex, added scope parameter, two-phase search, duplicate detection
  • Execution sandbox — persistent bash session, session-scoped tmp routing, write policy enforcement
  • Sub-agents — depth tracking, message queue, live state, timers, TODO anchoring
  • Token budget — soft/hard limits with graceful degradation
  • Tool system — entry-point plugins, MCP client, dynamic tool discovery

License

MIT. See LICENSE.