██████ ██████ ██████ ██████ █████ ██████ ██████ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██████ ██ ██████ █████ ██ ██████ █████ ██ █ ██ ██ ██ ██ ██ ██ ██ ██ ██ ███████ ██ ██████ ██ ██████ █████ ██ ██ ██████ ██ ██
A self-learning, multi-repo agent crew for Claude Code — and Cursor. Hand it one feature; it ships across every repo that feature touches — engineering its own context and learning your platform, so every run starts smarter than the last.
Simulate run · Install · First feature · What do you want to do? · Agents · Cheatsheet
Not a faster one-shot agent — a crew that fans out across your repos, engineers its own context, and gets sharper every run. One feature in, PRs across every repo out.
Nothing a one-shot agent learns survives the session. Your platform's conventions, the gotchas, the way you always do it — re-explained every run, like onboarding a new hire on a loop. And the moment a feature spans more than one repo, the agent that "finished" the backend has no idea the frontend and the contract drifted out from under it.
PipeCrew fixes both — with three moves:
- Multi-repo, multi-agent — one stack-specialized implementer and reviewer per repo, each in its own git worktree, all building against the same contract. A cross-repo assessor reads every diff together and catches the mismatch no single reviewer can see.
- Context engineered — state lives in files, not the chat. Each agent loads only its slice; the platform map, conventions, and decisions live on disk and are read on demand.
- Continuous learning — feed a merged PR back and PipeCrew proposes updates to its durable memory, which you approve per finding. Run #2 beats run #1.
You stay the director, approving at gates. The orchestrator job moves to PipeCrew.
PipeCrew ships its own demo. One command generates a complete fake workspace with realistic run artifacts and opens the live dashboard — the crew queuing, building, finishing — at zero agent cost:
claude plugin install pipecrew --marketplace pipecrew-ai/pipecrew
/simulate-runSixty seconds, no tokens spent, and you've seen the whole pipeline before onboarding a single repo.
Claude Code:
claude plugin marketplace add pipecrew-ai/pipecrew
claude plugin install pipecrew@pipecrewOr in one command (Claude Code v2.1.275+): claude plugin install pipecrew --marketplace pipecrew-ai/pipecrew
Cursor (v2.5+):
# Marketplace / git install (recommended):
# Customize → Plugins → /add-plugin → paste the repo URL
# github.com/pipecrew-ai/pipecrew
# Local dev against a clone:
cursor-agent --plugin-dir /path/to/pipecrewSame repo, dual-target — all 20 skills and the 35-agent crew ship to both; a few lifecycle hooks are Claude-Code-only for now (details).
Updates ship as GitHub Releases — enable
auto-update once via /plugin → Marketplaces → pipecrew, or see updating.
1. Onboard — once per project, from the directory that holds your repos:
/discover /path/to/your/reposScans the repos, detects tech stacks, asks a few domain questions, and writes the durable layer
beside your code: workspace config, a platform.md map of your domain, per-repo AGENTS.md
(the tool-agnostic standard — Claude Code, Cursor, Codex and 30+ agents read it natively), and
domain-specialized agents. Not sure what to build yet? Start with /brainstorm instead.
2. Deliver:
/deliver "publishers can choose contract type"Eight phases run automatically — requirements → architecture → contracts & specs land before
any code is written → plan → parallel build (backend + frontend + mock + infra, each in its own
worktree) → per-repo review → cross-repo assessment → report + draft PRs. You approve at each
gate; a live dashboard at localhost:5173 shows the crew work in real time.
Full phase-by-phase detail: cheatsheet.
3. Close the loop — after the PR merges:
/learn --pr <url>PipeCrew distills what reviewers changed into durable conventions — tier-classified, approved by you per finding — so the next run doesn't repeat the mistake. This is the pillar the other two exist for.
| You want… | Run |
|---|---|
| See the whole thing on a demo, free | /simulate-run |
| Ideas — what to build next, or how to build it | /brainstorm (--technical for approaches) |
| A greenfield project scaffolded from an idea | /scaffold |
| Your project onboarded | /discover |
| A feature shipped end-to-end across repos | /deliver "<description>" |
| A small, well-specified mechanical change | /patch |
| A branch or PR reviewed in one repo | /review |
| A feature branch verified across repos | /assess |
| Acceptance tests authored into a durable suite | /design-tests |
| The suite run — regression, UAT sign-off, prod smoke | /run-regression |
| A bug investigated from a symptom (read-only) | /troubleshoot "<symptom>" |
| Any concept, flow, or code explained — product or technical, with sources | /explain "<question>" |
| Feedback captured into durable conventions | /learn |
| Context docs audited or refreshed after drift | /context-refresh |
| Architecture diagrams generated or refreshed | /draw-diagram |
| Team memory shared via a private GitHub repo | /memory-sync enable |
| A teammate onboarded in one command | /join <memory-repo-url> |
| The current run, watched live | /site-view |
| Every Claude Code session on the machine, watched | /siteview-fleet (-list, -cleanup) |
Every row is a standalone skill — the pipeline is just the biggest one. Full reference with flags, agents, and supported stacks: docs/CHEATSHEET.md.
Behind the skills, a crew of 36 specialized agents. The ones you'll meet most:
| Agent | What it does |
|---|---|
solution-architect |
Designs the cross-repo solution — endpoints, schemas, boundaries |
security-consultant |
Flags auth gaps, data exposure, and injection in designs and diffs |
ux-consultant |
Produces an implementation-ready UX spec before the frontend is built |
architecture-mapper |
Infers cross-repo topology and draws the Mermaid diagrams |
context-manager |
Generates and refreshes AGENTS.md and agent-context docs |
feedback-learner |
Proposes tier-classified durable-context updates from a merged PR or run |
reporter |
Run report — waterfall timeline, per-agent tokens, real dollar cost |
Plus a paired implementer + reviewer for every supported stack (Spring Boot, React, Next.js,
NestJS, FastAPI, Flask, Django, Python workers, CDK, Terraform, mocks, schemas) — and the
domain-specialized product-owner / assessor / troubleshooter that /discover generates per
workspace. Full roster in docs/CHEATSHEET.md.
- Cheatsheet — every skill, flag, agent, stack, and phase on one page
- CHANGELOG · pipecrew.ai
Apache 2.0 · Learn more at pipecrew.ai