Switchman

You ran the agents. Did they build the right thing?

Is this for you?

✓ You run Claude Code, Cursor, or Codex in parallel worktrees
✓ You want to know agents built compatible code before merging
✓ You've had a "compiled fine, broke in prod" moment from parallel agents

If that's you, read on. If you're running a single agent on a single branch, you don't need this yet.

What it does

switchman start "your goal"
# agents run in parallel
switchman review

switchman review reads every worktree, turns a parallel session into a plain-English summary, flags semantic overlap and interface mismatches, and gives an honest merge confidence outcome: green, amber, red, or uncertain. When it cannot make a trustworthy call, it says uncertain instead of pretending everything is fine.

Built for developers using Claude Code, Cursor, Windsurf, Aider, and other CLI-first coding agents on real repos.

Questions or feedback? Discord · [email protected]

Install

Requirements: Node.js 22.5+ · Git 2.5+

npm install -g switchman-dev

Homebrew:

brew install switchman-dev/tap/switchman-dev

Switchman uses the built-in node:sqlite runtime — no extra database to install or manage.

Try it in 2 minutes

switchman demo

Creates a throwaway repo and shows:

• agent1 claiming src/auth.js
• agent2 getting blocked from the same file and rerouting safely
• both branches landing cleanly through the queue
• a final switchman review session summary

Then inspect it:

cd /tmp/switchman-demo-...
switchman status
switchman review

Quickstart

cd my-project
switchman quickcheck          # readiness check — prints one exact next command
switchman start "your goal"   # plans the work, creates workspaces, wires MCP
switchman status --watch      # live view while agents run
switchman review              # what was built, what doesn't fit, safe to ship?

When the review looks good:

switchman gate ci
switchman queue run

What switchman start sets up:

• repo context-aware task planning
• linked agent workspaces, one per agent
• shared Switchman database in .switchman/
• repo-aware CLAUDE.md when one does not exist
• automatic MCP coordination wiring where supported

If you want full manual control instead:

switchman setup --agents 3
switchman task add "Implement auth helper" --priority 9

Editor setup guides:

• Claude Code
• Cursor
• Windsurf
• Aider and Cline

The problem in detail

Three agents finish. You have three worktrees full of diffs and no coherent picture of what was built.

Git tells you what changed. It does not tell you whether parallel agent work is coherent.

The three things developers consistently hit:

• Review tax — 20 minutes manually reading diffs after every parallel session to piece together what actually happened
• Agentic drift — two agents independently solve the same problem in incompatible ways. No conflict. No CI failure. Just divergence that compiles, passes, and breaks later.
• No trustworthy answer — "is this safe to merge?" has been yours to answer alone until switchman review

Git branches and worktrees solve isolation. They do not solve coordination, ownership, visibility, or the question of whether a session produced something coherent. That is what Switchman adds.

How it works

1. Start with one goal

switchman start "Add user authentication" reads the repo, creates agent workspaces, and keeps work aligned as agents run. File claims, task ownership, and stale recovery happen automatically through MCP — you do not need to prompt agents to coordinate manually.

2. Review what was built

switchman review produces three things:

• A plain-English narrative of what each agent built — not a file list, a description
• Semantic flags — places where agents produced contradictory interfaces, duplicate implementations, or overlapping solutions
• A merge confidence outcome: green, amber, red, or uncertain

3. Ship through gates

When the review looks good, switchman gate ci and switchman queue run land finished work cleanly onto main.

Open source

Switchman is now one open source product. No login is required for the local workflow, and the core operator commands stay available out of the box.

switchman start "goal"   →   agents run   →   switchman review

What you get:

• switchman start — reads the repo, creates workspaces, writes MCP config, generates CLAUDE.md
• switchman review — full session summary, semantic overlap detection, interface mismatch flagging, merge confidence outcome
• switchman status --watch — live dashboard during a session
• switchman scan — lightweight pre-merge conflict check
• switchman demo — proves the flow in a throwaway repo
• 90-day searchable session history — switchman review --history, switchman review --history --search auth
• Cross-session pattern detection — switchman insights
• Shared review publishing — switchman review --share to publish, switchman review --team to read teammate reviews before the PR
• AI task planning — switchman plan "Add authentication" --apply, switchman plan --issue 47
• Cost and token tracking — switchman usage, switchman usage --days 30

Troubleshooting

switchman status          # main dashboard
switchman status --watch  # live view
switchman scan            # conflict scan across worktrees
switchman gate ci         # repo gate check
switchman verify-setup    # check MCP wiring

Explain commands for when something is blocked:

switchman explain claim src/auth/login.js
switchman explain queue <item-id>

More help:

• Status and recovery
• Merge queue
• Pipelines and PRs
• Command reference

PR gate

Block risky changes in GitHub the same way your local terminal does:

switchman gate install-ci

Installs .github/workflows/switchman-gate.yml — runs the repo gate on every push and PR.

What's next

• Zero-argument switchman plan — reads full repo context automatically
• Broader install and release distribution

Feedback

Building this in public. If you hit something broken or missing, I'd love to hear about it.

• GitHub Issues
• [email protected]
• Discord

License

MIT