angreal mcp advertises no tools, resources or prompts — those lists come back empty — and instead sends the task tree as the server's instructions. The assistant reads it, picks the task that fits, and runs it through the shell, so your own commands stay the entry point.
A Model Context Protocol server built into Angreal and started with angreal mcp. The client launches it as a stdio subprocess, and it answers initialize by returning the project's task tree — each task's description and risk_level — as the server's instructions.
- Your task tree delivered as system-level context, so the assistant knows which tasks exist before it reaches for a shell
- Per-task guidance from angreal.ToolDescription: when to use a task, when not to, and its risk_level
- A preview of exactly what the assistant will receive, printed by angreal tree --long
- JSON-RPC 2.0 over stdio, started and stopped by the client rather than run by hand
Angreal installed and on your PATH, and a directory holding an .angreal/ folder with task_*.py files — the mcp command is project-only. Configure the client with command angreal and args ["mcp"]; Claude Desktop's config is global, so point its working directory at the project.
One command — cargo install angreal_mcp