Shrimp Task Manager is an open-source MCP server with 2,100+ stars on GitHub that turns natural-language requests into structured development tasks for AI coding agents, with dependency tracking, reflection and iterative refinement built in.
Large coding tasks fail in agents for a predictable reason: the model tries to hold the whole plan in its head and loses the thread. Shrimp Task Manager externalises the plan. Running as a Model Context Protocol server inside editors and agent hosts such as Cursor, Windsurf and Roo Code, it takes a free-form instruction and decomposes it into an ordered task list with explicit dependencies, then walks the agent through the list step by step. The design emphasises chain-of-thought and self-reflection: tasks can be revisited, split and refined as the agent learns more about the codebase, which keeps long sessions coherent instead of drifting. Style-consistency guidance helps the generated work match the project it lands in. The project is MIT licensed; its last major updates landed in 2025, and it remains a popular reference for how to give reasoning models a durable working memory for project-shaped work.
Shrimp Task Manager is free and open source under the MIT license. You install the server locally and use it with your own AI editor or agent; any model costs belong to the host tool you pair it with.
Developers using AI coding agents on multi-step projects who find that the agent loses track of the plan, and who want a structured task layer that keeps big refactors and feature builds organised from start to finish.