open source · Claude Code, Cursor, Codex · v0.4.1

Ran AI agents
in parallel?
Is it safe to merge?

Run your agents however you want. When they're done, one command tells you if it's safe to merge — before anything touches main.

● GREEN
● AMBER
● RED
← your answer, in seconds
You run Claude Code, Cursor, or Codex in parallel worktrees
You want to catch semantic drift before it reaches main
You've had a "compiled fine, broke in prod" moment from parallel agents
$ npm install -g switchman-dev

Open source, MIT licensed. No setup required — use it after your agents finish, before you merge.

What every parallel agent session ends with
Three agents finish. You have three worktrees full of diffs and no idea if any of it fits together.
No clear picture
Agents are fast. The manual review at the end is the bottleneck that kills the time saving.
Agent 1's middleware expects req.user.id. Agent 2's schema has no id field — only userId.
Agentic drift
No merge conflict. No CI failure. Just a mismatch that compiles, passes tests, and breaks in production.
You stare at four branches and ask yourself the question git can never answer for you.
"Is this safe to merge?"
Git tells you what changed. Nothing tells you whether it's coherent. That's been your job — until now.
How it works

Run your agents. Then run Switchman. That's it.

No setup before you start. No changes to how you work. Just run it when your agents are done and before you merge.

01

Run your agents however you want

Claude Code, Cursor, Codex, Aider — in separate worktrees, branches, or directories. Switchman doesn't care how you ran them or what tools you used.

02

Run one command when they finish

switchman review --pr-ready --all-worktrees

Switchman scans every branch, reads what each agent built, and checks for semantic overlap and interface drift.

03

Get a clear answer before you merge

Green, amber, or red — with a plain-English PR-ready report showing exactly what was built, what overlapped, and what needs a human eye before it touches main.

$ switchman review --pr-ready --all-worktrees
# scan all worktrees after agents finish
$ switchman review --pr-ready --from branch-a branch-b
# review specific branches
$ switchman gate ci
# CI gate — fails if confidence is red

Try it in 2 minutes — no repo needed

Creates a throwaway repo, runs two agents that conflict, and shows you an amber catch in real time. The fastest way to see what Switchman actually does.

$ switchman demo
Auto-trigger

Runs without thinking about it.

Install once, forget about it. Switchman fires automatically when your agents finish — no manual step, no context switch, no forgetting.

Watch mode — any agent

Scan when worktrees go quiet

Polls all worktrees continuously. When activity stops for 5 seconds, a scan fires automatically. Works with any agent — no hook support required.

$ switchman watch
# or run as a background daemon:
$ switchman monitor start
CI & PR integration

The report lives where review already happens.

One command drops a GitHub Actions workflow that posts the merge confidence result as a PR comment — visible to your whole team, not just you.

$ switchman gate install-ci
# drops .github/workflows/switchman-gate.yml — done
🤖
switchman-bot commented on this pull request ● AMBER

Switchman Merge Confidence

🟡 Amber — Review before merging. Parallel agent changes detected.

SignalValue
Merge confidenceamber
Gate statusblocked
AI gatewarn
Non-compliant worktrees1
Stale worktrees0
Review Signals
Semantic conflicts: 1
Ownership conflicts: 1
Review flagged worktrees locally with switchman review --pr-ready --all-worktrees before merging.
⭐ Switchman caught a risky merge? Star us on GitHub — the ask only appears when it's earned it.
What Switchman catches

The things git diff can't tell you.

Merge conflicts are the easy problem. Switchman finds the hard ones — the semantic mismatches that compile, pass tests, and break in production.

The real risk isn't merge conflicts — it's agentic drift

Parallel agents independently solve the same problem in incompatible ways. No conflict. No CI failure. Just divergence that compiles, passes, and breaks things later. Switchman detects semantic mismatches before they reach main — the thing git diff simply can't tell you.

Core command

switchman review

Plain-English summary of what each agent built, semantic overlap flags, interface mismatches, and a green / amber / red merge confidence score. Generates a PR-ready report in 30 seconds.

Honesty

Four confidence states

Green, amber, red — and uncertain. When Switchman can't determine coherence it says so. A tool that admits uncertainty is one you can trust when it says green.

Detection

Semantic overlap

Finds cases where two agents solved the same problem differently — incompatible interfaces, divergent data models, duplicate logic that will collide at runtime.

CI integration

switchman gate ci

Blocks merges when confidence is red. One command installs a GitHub Actions workflow. Posts the result as a PR comment visible to your whole team.

Coordination (optional)

Parallel agent orchestration

For teams that want more: file locking, task queues, lease management, and conflict prevention via MCP. Agents coordinate automatically — no prompting required.

Visibility

Live session dashboard

switchman status --watch — what's running, blocked, or done across all agents. The thing developers leave open on a second screen.

Works with

Use whatever agents you already use.

No changes to your workflow. Run Claude Code, Cursor, Codex, or Aider however you normally would — then run Switchman when they're done.

Windsurf MCP install
Aider CLI
Cline CLI
Works with any agent that writes to a git worktree or branch

Ran agents in parallel?
Find out if it's safe to merge.

Open source · MIT licensed · Free forever