Quickstart — Claude Code

Pome gives your agent a graded exam: a clone of it (the examinee) works a real task against deterministic digital twins of GitHub and Slack, every tool call is recorded, and a judge scores the tape. The agent driving the test (the coach — your Claude Code session) installs, connects, runs, and brings the failure summary back into the conversation — hands off.

Already have your own agent in a repo? Run this example first to learn the loop, then see Bring your own agent to register it and get a pome.json.

Before you start

Claude Code installed and signed in. The example examinee is a small Claude Agent SDK process, so your stored claude login (or ANTHROPIC_API_KEY) covers its auth too.
Node 18+ with npm and npx.
No Pome account yet? You create one inside the OAuth consent flow in step 2 — no credit card.

1. Install the coach skills

npx skills add pome-sh/digital-twins

The skills CLI installs the Pome coach skill set from pome-sh/digital-twins into your agent’s skills directory: pome (the entry router), pome-intake, pome-author-task, pome-verify-seed, and pome-run-task. They teach Claude Code the full loop — including the one rule that loses runs when missed: finalize while the twin session is still live.

2. Connect the Pome MCP

claude mcp add --transport http pome https://mcp.pome.sh/mcp

Every command here is plain text you can read before you run it — the install path never pipes a script into your shell and never injects remote config. There is no API key to paste: on the first tool call, your agent discovers the OAuth metadata and opens the Clerk consent screen (scopes: verify your identity, basic profile, email address). Sign in — or sign up right there — and the tools appear.

3. Paste the prompt

Open Claude Code anywhere (the example brings its own agent — your repo is not required) and paste:

Test my agent with pome, using the bundled support-triage example, and take it
all the way to a finalized, scored run — hands off.

First, present me this plan as a short checklist and ask "Shall I proceed?"
before running anything. Wait for my yes.

1. Fetch the example pack and build the local examinee:
     npx degit pome-sh/digital-twins/examples/support-triage support-triage
     cd support-triage/local && npm install
2. On the pome MCP, call get_started, then get_platform_context with
   door: "claude-code" to confirm identity and quota (the door tag is for
   onboarding analytics — pass it verbatim).
3. Register the examinee: register_agent with name "support-triage" and
   twins ["github", "slack"]. Keep the agent id.
4. Save the exam into my team catalog: save_task with the full contents of
   ../tasks/duplicate-issue.md as task_source and the name
   "support-triage-dedup". Keep the task id.
5. run_task(task_id, agent_id). Keep session_id, agent_token, examinee_task,
   and examinee_launch from the response.
6. Spawn the examinee as a local subprocess: from support-triage/local, run
   `npm run start` with env injected from the run_task response —
   POME_GITHUB_MCP_URL and POME_SLACK_MCP_URL from the github and slack
   entries' url in examinee_launch.mcp_servers (not examinee_launch.env,
   whose URLs are REST), POME_AUTH_TOKEN = agent_token,
   POME_TASK = examinee_task.prompt. POME_AUTH_TOKEN is env-only: never
   write it to a file.
7. The instant the process exits, call finalize_run(session_id, agent_token)
   — before any cleanup, while the twin session is still live.
8. get_report(run_id). Give me the score, the criteria table, and a plain
   summary of what failed and why, here in the conversation, plus the
   dashboard link.

Guardrails: pause after the checklist for my go-ahead; keep every command plain
text and show it to me before running; never pipe remote scripts into a shell;
secrets stay in env vars; don't ask me for an API key — the OAuth consent
screen handles auth.

Claude Code does the rest: it seeds live GitHub + Slack twins, launches the examinee against them, waits for it to finish, finalizes, and narrates the report.

The twin tape lives in the running session. finalize_run must be called while the session is still live — the installed pome-run-task skill enforces this, and the prompt repeats it because it is the one unrecoverable mistake.

What the score means

Expect 33, not 0 — and not 100. The task is a support-triage trap: a customer re-reports a bug that open issue #1 already tracks. The bundled v1 examinee does real work (its report has concrete repro steps — that’s the 33) but files a duplicate issue instead of finding the existing one, flunking both duplicate-detection criteria. The report shows exactly which criterion failed and what the judge saw — the failure a happy-path demo never surfaces.

The one-line fix

The whole story is one line of the examinee’s system prompt: TRIAGE_RULE in support-triage/local/src/index.ts. Swap it for the commented TRIAGE_RULE_V2 right next to it (search open issues first; comment on the existing issue instead of opening a second one), then tell Claude Code:

Apply the TRIAGE_RULE_V2 fix in support-triage/local/src/index.ts and re-run
the same task (steps 5–8). Show me the score delta.

Same exam, same twins, green run. That loop — run, read the failure, fix, re-run — is the product.

Open your first report

get_report prints the score, the criteria table, and the dashboard link right in your conversation. Open the link to see the finalized report — a failing run where the “33, not 0” story is legible: which criterion failed, and exactly what the judge saw.

A finalized Pome report scoring 33 out of 100, with the two failed duplicate-detection criteria highlighted. — A finalized support-triage report — 33/100, with the failing criteria named.

Stuck before your own run lands? Look at a real one — no signup: the failing baseline and the same task after the one-line fix.

MCP workflow

The coach loop behind the paste-prompt, call by call.

Bring your own agent

Point Pome at your own repo and get a scored run.

Author your own task

validate_task → save_task → verify_seed — graded exams for your agent.

Dashboard

Where runs, traces, and judge feedback live: app.pome.sh.

Get started

Guides

Command Line Interface

MCP

Quickstart — Claude Code

Before you start

1. Install the coach skills

2. Connect the Pome MCP

3. Paste the prompt

What the score means

The one-line fix

Open your first report

Next

MCP workflow

Bring your own agent

Author your own task

Dashboard

​Before you start

​1. Install the coach skills

​2. Connect the Pome MCP

​3. Paste the prompt

​What the score means

​The one-line fix

​Open your first report

​Next

MCP workflow

Bring your own agent

Author your own task

Dashboard

Before you start

1. Install the coach skills

2. Connect the Pome MCP

3. Paste the prompt

What the score means

The one-line fix

Open your first report

Next