Wires TangleBrain's tanglebrain-delegate stdio
MCP server into Claude Code, so the orchestrator can offload self-contained sub-tasks — code
generation, drafting, extraction, summarization, bulk transforms — to a free local model (or
any configured can_delegate backend) and keep its own budget for decomposition and review.
Exposes four tools: delegate_local, delegate, delegate_many, delegate_targets. See the
project README
for the full tool semantics and routing rules.
The plugin registers the server declaratively — it does not vendor the Python code. Install
TangleBrain with the delegate extra so the tanglebrain-delegate command is on your PATH:
pip install "tanglebrain[delegate]"The server reads the roster from $TANGLEBRAIN_ROSTER → ~/.config/tanglebrain/roster.yaml → the
packaged example, in that order; copy the example to the user-config path and point it at your
backends.
/plugin marketplace add Jason-Vaughan/TangleBrain
/plugin install tanglebrain-delegate@tanglebrain
Equivalent manual registration (no plugin): claude mcp add tanglebrain-delegate -- tanglebrain-delegate.
If the prerequisite is missing, the server shows as failed in /mcp (Claude Code spawns
tanglebrain-delegate and gets command-not-found) — the plugin itself installs fine either way.
Check which tanglebrain-delegate in the same environment Claude Code runs from; if it's absent,
run the pip install above. If the command exists but tools error at call time, the roster is the
next suspect: run tanglebrain-delegate by hand and call delegate_targets to see the live error.