Task lists are easy to build and this one does not pretend otherwise — what it adds is the step where a vague task becomes a set of tractable subtasks with a complexity estimate attached, and where a product document becomes a plan without someone transcribing it. The multi-system bridging is worth noting if you already use another task server: this can sit in front rather than asking you to migrate.
A Python task-management system with an MCP interface, built on LangGraph's ReAct pattern. Beyond create-list-update-delete it does the part that is normally manual: splitting a large task into subtasks, judging complexity, and turning a PRD into structured work. It can also connect to other task systems rather than replacing them.
- Task management with status tracking and dependency handling
- Task decomposition — break a complex task into subtasks, with complexity assessment and automatic PRD parsing
- Model choice rather than a fixed provider: OpenAI, Claude and others through the same interface
- An interactive command line as well as the MCP server, so the same task store is reachable both ways
- Bridges to existing task systems including mcp-shrimp-task-manager and claude-task-master
- A LangGraph Studio path for visualising the agent graph, stepping through execution with breakpoints and editing state mid-run
At least one model key — `OPENAI_API_KEY` or `ANTHROPIC_API_KEY` — in a `.env` file at the project root. Optional model settings there too: `LLM_MODEL` defaults to `gpt-4o`, `TEMPERATURE` to 0.2 and `MAX_TOKENS` to 4000. Install with `uv pip install -e .` or `pip install -e .`, plus `npm install` for the MCP server's Node dependencies. The MCP entry point is `run_mcp.py`, launched over stdio with the model key in the client's env block. LangGraph Studio, if you want it, needs `langgraph-cli` 0.1.55 or higher and Python 3.11+. MIT licensed.
One command plus a key — uv pip install -e ., then supply credentials