Open source Beta

Stop trusting coding agents when they say they're done.

Your coding assistant plans. Sandboxed workers build. Nothing counts until nomArmy has checked it.

AI workers' “done, all tests pass” is a claim, not evidence. nomArmy lets your coding assistant hand work to workers called noms, then checks every change itself before anything is committed.

For Claude Code, Codex and Cursor. On macOS, Linux and Windows via WSL2.

Measured in our own use

About 1 in 11 “done” claims wasn't true.

265
times a worker said “done, tests pass”
24
weren't true: 16 failed when nomArmy ran the tests itself, and 8 had tests that still passed with the change reverted
140
new test files shown to fail without their change
33 of 38
high-stakes changes reviewed by a model from another vendor

From nomArmy's verified job records on one team's machine, September 18 to October 4, 2026, across nomArmy itself and our own products. One team's experience, not a benchmark. How we measure.

A clear division of responsibility

How it works

  1. The General plans

    Your coding assistant briefs the task, acceptance criteria and tests that prove it.

  2. Sandboxed noms build

    Workers implement in isolated worktrees and Podman sandboxes, without network access or host credentials.

  3. nomArmy verifies

    It reads the real diff, runs tests in a fresh sandbox, reverts a production change to check the tests catch it, and scans for secrets.

  4. You merge

    Verified work is committed on the worker's own branch. Reviewing and integrating stay with the General and you.

Claude subscription tools run on the host, so build jobs are refused unless explicitly allowed. Read the security posture.

Work you do not have to take on faith

What's different

Independently verified results

A worker's report does not decide the outcome. Failing verification stays failed, and tests that pass without their change go to review.

Promises kept in the repo

Each feature's acceptance contract lives beside its code, with the tests that prove every criterion. A change that breaks or quietly weakens a promise is flagged before it's committed.

How feature runs use them

Trust gates for sensitive code

Paths you mark sensitive, removed or bypassed access checks and tenant filters, and an optional model judgment raise a change to review by another vendor, or to a hard stop until you sign off.

Map your trust boundaries

Your tools. Your models.

Works with Claude Code, Codex and Cursor. Use local models, API keys or your own subscriptions. Run on macOS, Linux or Windows via WSL2. Even the AI trust judgment can run on a server on your own machine, so your code never has to leave it. Measure a local model first.

Find your setup

Start with one small task

Give your General an army.

Have Git, Node 24.16+ (or 26.1+) and Podman ready. Install the CLI, open your project, then run the setup playbook.

Full installation guide

npm install -g nomarmy
cd your-project
nomarmy setup

Setup shows a checklist and runs the next step when you say yes. Stop anytime and pick up where you left off.

Using Claude Code or Codex? After setup, you can add nomArmy as a plugin instead: /plugin marketplace add rayson-tech/nomarmy. How

View nomArmy on npm

Evidence you can share

A claim is not a result.

Feature runs produce a “Verified by nomArmy” block for your pull request. It shows which claims held up, what was caught, and which new tests failed without their change.

Explore feature runs

nomArmy: 72 claims checked, 5 caught

Repository badge snapshot, not live totals.

âś“ Verified by nomArmy

Illustrative pull request example

Worker claims checked
2 of 3 “done, tests pass” claims held up; 1 caught when nomArmy ran the tests itself.
Tests proven
2 new test files shown to fail without their change.
Work committed
2 verified jobs, +20 / -4 lines.

In your PR, these numbers come from verified job records, not the workers' own reports.