An AI agent exploration. This is not a coherent product or framework. This is a laboratory in the shape of a Rube Goldberg machine. Built to explore agent development patterns. Increasingly focused on personal knowledge management and writing tools, with an Obsidian-like shared vault where user and agent collaborate on markdown documents.

Multi-channel AI agent with a shared knowledge vault. Connects to Mattermost as a chat bot, runs in a web UI with WYSIWYG wiki editing, or runs in terminal mode. Multi-provider LLM support (Vertex/Gemini, OpenAI, OpenAI-compatible) with named model configs and per-conversation model selection. Streams responses as they arrive.

See docs/ for the full feature list.

git clone https://github.com/lmorchard/decafclaw.git
cd decafclaw
uv sync
cp .env.example .env    # then configure LLM provider — see docs/providers.md
make run                 # interactive terminal mode (no Mattermost needed)

See Installation & Setup for provider configuration, Mattermost bot setup, and all options.

make dev       # Auto-restart on file changes
make test      # Run pytest
make check     # Lint + type check (Python + JS)
make vendor    # Rebuild web UI vendor bundle
make config    # Show resolved config values

Most major features are developed in dev sessions, with design artifacts preserved under docs/dev-sessions/ — each session directory holds a spec.md, plan.md, and notes.md capturing the thinking behind the change. Browse those for the history of how the project grew.

MIT

Do what you want with it, if you can make sense of it. I barely can, myself, but I'm learning. 😅