An agent that plans and implements in one breath has no moment where the plan can be corrected. Here the analysis and the reflection happen before any code, and verification happens after — with the task list on disk, so hitting a context limit costs you a session rather than the work.
A task layer for coding agents. Work is planned, analysed, reflected on and split into atomic units with their dependencies tracked; execution and verification are separate steps, and the whole set persists on disk so a new session resumes rather than restarts.
- A plan produced from a stated goal, then analysed for what it actually requires and reflected on before anything is written
- Large work decomposed into atomic, testable subtasks with the dependencies between them tracked
- Execution as its own call, followed by verification as another — so finished means checked rather than claimed
- Tasks listed by status, queried by criteria, opened in full detail, updated and deleted individually
- Every task cleared at once with a backup taken first, so a clean slate is not a loss
- Project rules initialised once and then applied, which is how coding standards survive across sessions
- A research mode for systematic exploration of a technology before committing to it
- A thought-processing step the agent uses to structure its own reasoning as part of the flow
- State on disk under a directory you choose, which is what makes the memory outlast the conversation
Node.js 18 or newer and an MCP-capable client. Installed from a clone: npm install, then npm run build, then point your client at the built dist/index.js. Three environment values shape it — DATA_DIR, which is where tasks and history are written and should sit inside the project you are working on, TEMPLATES_USE for the prompt language, and ENABLE_GUI, off by default, which turns on a lightweight web overview. A separate React task viewer ships in the repository and runs on its own.
One command — npx -y mcp-shrimp-task-manager