# pawbot

pawbot is a self-hosted AI agent reachable from a browser, a terminal or chat apps, that reads and writes files, runs commands, searches the web and calls MCP tools, with a Record and Replay feature that replays a past turn offline while the code changes.

A user chooses their interface: a visual WebUI workspace, a native terminal UI for speed, or a connected chat app for reaching the agent from anywhere, all backed by the same session history and long-term memory across conversations. Record and Replay captures one real agent turn and lets a developer replay it later without another model request, another token charge or real tool side effects, which the project positions as useful for verifying an AI-assisted code change or for offline evaluation.

Memory can be reviewed and adjusted in a personalization settings area, explicit preferences can be saved immediately with a remember command independent of automatic memory compaction, and natural-language remember requests are turned into reviewable candidates rather than applied immediately. The agent supports Anthropic, OpenAI-compatible endpoints, local models and configurable fallbacks, and exposes a Python interface alongside its WebUI, CLI and API.

It is written in Python, with English and Chinese documentation.

Website: https://github.com/m2dumpling/pawbot