Open Source

Claude PM Toolkit

A project management intelligence layer that transforms AI coding agents from ad-hoc code generators into disciplined engineering partners. Structured workflows, adversarial reviews, parallel development, and persistent institutional memory.

By the Numbers
49
MCP intelligence tools
4
Core workflow skills
2 min
Install time
MIT
License
The Problem
AI coding agents without discipline create chaos
  • Duplicate issues for the same problem. Cluttered backlog, wasted effort.
  • Bundled unrelated changes into single PRs. Unmergeable pull requests, blocked reviews.
  • "LGTM" rubber-stamp reviews. Missed bugs and unhandled edge cases ship to production.
  • Lost context between sessions. Rework, repeated questions, starts over from scratch every time.
  • No workflow discipline. No visibility into what's in progress, what's blocked, or what's done.
Four Core Skills
Slash commands that enforce engineering rigor

/start

Session kickoff with parallel intelligence gathering: risk radar, anomaly detection, workflow health, stale decisions, and an optimized work plan. Supports time constraints and area focus.

/issue

Full issue lifecycle. Create mode runs a PM interview with AI triage and duplicate scanning. Execute mode loads full context with predictive intelligence and enforces scope discipline.

/pm-review

Adversarial code review with blast radius modeling, rework prediction, mandatory failure mode analysis, and a learning system that tracks false positive rates over time.

/weekly

AI-generated weekly narrative with DORA metrics rated against industry benchmarks, Monte Carlo sprint forecasts, risk radar, and stakeholder-friendly framing.

Intelligence Layer
49 MCP tools for project intelligence
The toolkit exposes a Model Context Protocol server that gives AI agents direct programmatic access to project intelligence. Not just status queries: prediction, simulation, and learning.

Prediction

P50/P80/P95 completion forecasting, rework probability with weighted signals, and DORA metrics tracking.

Simulation

Monte Carlo sprint throughput, backlog forecasting, and cascade modeling for slip scenarios.

Memory

Persistent decision logging, outcome tracking, rework rate analytics, and cross-session event streams. Nothing is forgotten.

Guardrails

Scope creep detection, AI context waste metrics, WIP limit enforcement, and workflow health monitoring.

Parallel Development
Multiple issues, simultaneously, without conflicts
Each issue gets its own Git worktree with isolated ports. A tmux portfolio mode runs multiple AI coding sessions simultaneously in separate windows with a status bar that alerts when a session needs input. Multiply your throughput without multiplying your headcount.
Architecture
Local-first, framework-agnostic
  • GitHub stays source of truth for issue content, PRs, labels, and CI. Local SQLite owns workflow state, dependencies, and analytics.
  • Sub-millisecond board queries. No API rate limits. Works offline. Richer data modeling than GitHub Projects.
  • Strict Kanban enforcement. Backlog, Ready, Active, Review, Rework, Done. WIP limit of 1. Every state transition recorded.
  • Installs in 2 minutes. Auto-detects your stack. No vendor lock-in. Works with any Git repository.
Discuss a build like this