Skip to main content
Pome runs your coding agent against a deterministic clone of a real SaaS API—GitHub, Stripe, or Slack—recording every tool call and scoring the result. You drive it entirely from the agent you already use. The CLI ships two skills that wire up your agent and carry it through a first scored run in minutes.

​
1. Install

npm install -g pome-sh
pome version
Stage 1 pre-launch: pome-sh is not yet on npm. Until the launch publish, install from a local checkout: git clone https://github.com/pome-sh/pome && cd pome/cli && bun install && bun run build && npm install -g . --ignore-scripts.

​
2. Log in

pome login
A browser opens at app.pome.sh, you sign in, and a pme_… API key is stored in your macOS Keychain (preferred) or ~/.pome/credentials.json (Linux/Windows or when Keychain is unavailable). CI can skip this and set POME_API_KEY directly.

​
3. Install the Pome skills

The Pome skills teach your coding agent how to wire itself up to pome and run scenarios. 
pome skills install
Two skills land in your agent’s skills directory (~/.claude/skills/, symlinked to the installed pome-sh package so npm i -g pome-sh@latest keeps them current):
  • /pome-setup: identifies your agent and the services it talks to, registers it on the dashboard, and writes a starter TESTS.md.
  • /pome-test: runs the scenarios that match your agent and reports the results.

​
4. First run

Open your coding agent (Claude Code, for example) in the project that contains your agent code. Then prompt it: 
set up my agent to test with pome. Use /pome-setup
/pome-setup will:
  1. Identify your agent + the services it uses (GitHub? Stripe?) and confirm with you.
  2. Add minimal, non-breaking hooks so pome can capture tool calls during a test run.
  3. Register the agent on the dashboard and print a URL.
  4. Suggest 5 scenarios from the library that match your agent and ask for confirmation.
Once you confirm, prompt the agent again: 
test my agents with pome
/pome-test runs the confirmed scenarios against the matching twin and reports back inline.

​
5. See the results

Open the dashboard: 
https://app.pome.sh
Each run shows the trace, the score, and the judge feedback for any criterion that failed. Local copies of the artifacts (events.jsonl, score.json, state snapshots) stay on disk under runs/<scenario>/<run-id>/.

​
Run the twin locally

pome run defaults to the hosted twin on app.pome.sh. To run the twin yourself, pull the image and start it with Docker: 
docker run --rm -p 3333:3333 ghcr.io/pome-sh/twin-github
The image exposes port 3333. Then point pome at the local twin with POME_LOCAL=1: 
POME_LOCAL=1 pome run scenarios/<your-scenario>.md
Until the Stage 1 public flip, the GHCR image is private. Run docker login ghcr.io first.
The standalone twin path is documented in the GitHub twin guide.

​
Next

/setup

Wire your agent with /pome-setup.

/test-with-pome

Run scenarios with /pome-test.

GitHub twin

Twin reference and bundled GitHub scenarios.

CLI reference

Command reference: pome run, pome session, flags, and exit codes.

Dashboard

Where runs, agents, and twins live, and how the judge surfaces fixes.