Coding agents are great at writing a single function and terrible at holding a multi-step plan in their head. Shrimp's bet is that the missing piece isn't a smarter model but an external task graph the agent is forced to consult: requirements become atomic tasks with explicit dependencies, and the agent must plan, reflect, and check off each node before moving on — so progress survives context resets instead of evaporating when the conversation scrolls.
What Sets It Apart
- Tasks are a dependency graph, not a flat checklist: the server tracks which units block which, so an agent can't skip ahead to work whose prerequisites aren't done — useful when one prompt would otherwise produce half-finished, out-of-order edits.
- Persistence is the real feature. Task state and progress live outside the chat, so a fresh session (or a crashed one) picks up exactly where the last left off — the usual failure mode where an agent forgets what it already built.
- A built-in research mode and reflection step push the agent to investigate options and critique its own plan before generating code, trading a few extra turns for fewer dead-end rewrites.
- Ships with a React task viewer so a human can watch the agent's plan unfold and intervene, rather than trusting an opaque autonomous run.
Who It's For
Great fit if you drive a coding agent (Claude Code, Cline, Claude Desktop) through large, multi-file features and keep losing the thread across sessions. Look elsewhere if your work is one-shot scripts or quick edits — the overhead of decomposing requirements into a tracked graph only pays off when the task is big enough to outlive a single context window, and it adds a planning layer some workflows won't want.
