Because tasks and memories are files under the project's working directory, the breakdown can be committed and reviewed like code, and the next session starts from what was agreed rather than from a summary. Two tools do work nothing else here does: infer_task_progress reads the codebase and proposes which tasks are already finished, which is how a stale board gets corrected, and get_next_task_recommendation answers "what now" from dependencies and complexity rather than from list order. Note that the subtask tools remain only for compatibility — new work belongs in the unified task tree.
Task management and persistent memory stored per project. Every tool takes a workingDirectory, and the data lands in that directory so it can be committed alongside the code — or in one global directory when the server is started with --claude.
- Projects created, listed, updated and deleted, each holding its own task tree
- Tasks nested to unlimited depth through parentId, with move_task relocating a branch and list_tasks rendering the hierarchy
- Every task carries priority and complexity on a 1–10 scale, dependencies on other tasks, tags, estimated and actual hours, and a status of pending, in-progress, blocked or done
- parse_prd turns a product requirements document into structured tasks, optionally generating subtasks and estimating complexity as it goes
- get_next_task_recommendation picks what to work on from dependencies, priorities and complexity, and can exclude anything currently blocked
- analyze_task_complexity flags tasks that should be broken up, suggests the breakdown, and can create the subtasks itself
- infer_task_progress reads the codebase — file creation, implementation evidence — and proposes which tasks look done, with a confidence threshold and optional automatic updates
- research_task guides a web research pass for a task and saves what it finds into memories; generate_research_queries produces the search queries to run
- Memories with title, content, category and metadata, searched across all three fields with relevance scoring and a threshold
- Deletions require an explicit confirm, and migrate_subtasks converts data from the older three-level model to the unified task tree
No account and no key — everything is stored on disk. It runs as npx -y @pimzino/agentic-tools-mcp, and data lands in a .agentic-tools-mcp directory inside whatever workingDirectory each call names. Starting it with --claude switches to one global directory instead (~/.agentic-tools-mcp on macOS and Linux), and in that mode the workingDirectory argument is ignored. A companion VS Code extension provides a GUI over the same data.
One command — npx -y @pimzino/agentic-tools-mcp