From vibe coding to Vibe Engineering.

The loop

Four beats, and the same four every time. What makes the result reliable is the loop, not how much software engineering the person driving it happens to carry in their head.

1. The requirement

   It starts where every task starts: one sentence describing what should be true when this is done. Nothing about the agent, the branch or the tooling yet.

   “shout() should return the greeting uppercased. npm run test:shout currently fails.”

2. The contract

   Before the work starts, the sentence is written down as a contract: what may be touched, and which gates decide that it is finished. Two files, both in the repository, both readable by a human.

   .commandmate/verify.yaml

   version: 1
   gates:
     - id: unit
       command: npm test
       timeoutSec: 120

   .commandmate/tasks/fix-shout.yaml

   title: "shout() uppercases the greeting"
   scope:
     allow: ["src/greet.js"]
     deny: ["test/**", ".commandmate/**"]
   verify:
     gates: [issue-shout]
   success:
     requireWorkEvidence: true
     requireScopeClean: true

3. The agent runs

   The contract is handed to whichever coding agent you want to use, in a Git worktree of its own. Several tasks can run at once because none of them shares a working directory, and you can watch any of them from a browser.

   $ commandmate send wt-shout \
       --contract .commandmate/tasks/fix-shout.yaml >task-id.txt
   Task created: 4d1f8b2a-6c07-4a5e-9f3b-27ac91e0d4c8
   Message sent.

4. The verdict

   When the agent stops, the gates you declared run for real and the process exit code is the answer — 0 passed, 20 a gate failed, 21 nothing was actually done. Not the agent's own opinion of its work.

   $ commandmate wait wt-shout --verify
   GATE work-evidence PASS (commits=0, uncommitted=1)
   GATE scope PASS (exit=0, 0.1s)
   GATE unit PASS (exit=0, 0.1s)
   RESULT passed
   $ echo $?
   0

What it gives you

Runs 100% locally. No external server, no cloud relay, no account required — the only network traffic is the agent CLI's own API calls.

Quick start, in 60 seconds

Two ways in. Take Track A to see it working, Track B once you keep coming back.

Either track leaves a server running in the background. commandmate status tells you whether one is up, and commandmate stop shuts it down.

Then run the loop on a real repository

Use a sample repository with two bugs left in on purpose to work through the core of CommandMate in about fifteen minutes. You fork the sample repository before you start, so nothing you do can touch the original repository (upstream). You hand the agent a contract before the work and let the verification gates judge it afterwards, and you read the verdict off the real exit code.

The sample repository — fork this one first:

Fork it on GitHub, then paste your fork's URL into Repositories → Add Repository → Clone URL and follow the tutorial. It ships its own .commandmate/verify.yaml and two task contracts, so the gates are real from the first command. No install, no dependencies.

See it running

Twenty seconds each, recorded against a throwaway repository with a stubbed agent. The gates in the first one are executed for real, so the exit codes on screen are the ones that run produced.

With and without CommandMate

The comparison worth drawing is between two ways of working, not between products.

On desktop and on mobile

The same sessions in your browser, whichever screen you are on.

Pair your phone with a QR code

One command opens a way in and prints a QR code carrying a one-time pairing code. There is no LAN address to look up, no CM_BIND to widen, and no token to retype on a phone keyboard.

Nothing is exposed until you run it, and the session closes itself — eight hours by default, --expires to change it. commandmate remote status reports the provider, the URL and what is left of both windows, and Auto-Yes stays off unless you turn it on.

The Quick Tunnel is a quick way in rather than a permanent address. Every flag, every exit code and the longer-lived Tailscale route are in the CLI operations guide.