> ## Documentation Index > Fetch the complete documentation index at: https://docs.pome.sh/llms.txt > Use this file to discover all available pages before exploring further. # Gmail twin in 5 minutes > Start a private mailbox holding an unanswered customer email, let your own coding agent reply and file it, then read the twin's recorded tape — including what it deliberately refuses to keep. No Google account, no OAuth consent, nothing graded. **About 5 minutes.** You start a private mailbox holding one unanswered customer email, your own coding agent replies to it and files it, and you read the twin's own record of every call it made — including what the recorder keeps about a sent message and what it deliberately throws away. A **digital twin** is not a mock. It is a stateful service that answers the same REST and MCP calls as `gmail.googleapis.com`, boots from a declared starting state, records every request, and never reaches Google. No mailbox is connected and no OAuth consent is asked for, because there is no real account behind it to consent to: the hosted twin holds no Google credential, ever. ## Before you start * **A Pome account.** `pome login` opens the browser sign-in and creates one if you do not have it. No credit card. * **Your own coding agent** — Claude Code, Cursor, anything that can run a shell command and read JSON back. * **Node 18+** for `npx`, plus `curl` and `jq` for the transcripts below. Your agent can read the raw JSON without `jq`; it is here to keep the blocks short. Nothing else. No Google account, no `ANTHROPIC_API_KEY`, no model inference paid for by Pome: **your agent is both the operator and the actor.** It drives Pome, and it is the thing that acts on the twin. **Nothing on this page is graded.** No task file, no criteria, no score — and no agent eval is charged, because an eval is only ever burned when a run is graded. A sandbox you start, drive and stop costs you nothing. Grading appears exactly once in this curriculum, at the [support-triage capstone](/quickstart/claude-code), where the agent under test is sealed off from the criteria that judge it. ## Paste this Hand this to your coding agent as-is. It names the twin's own Gmail surface and the boundary it must not cross. ```text theme={"dark"} Set up a Pome Gmail sandbox, answer the unread customer email in it, then show me the recorded tape. Show me this plan first and wait for my go-ahead. Keep every command plain text; never pipe a remote script into a shell. 1. Sign in and start the world: npx @pome-sh/cli@latest login npx @pome-sh/cli@latest sandbox create --twin gmail \ --secrets-file .pome-sandbox.env --format json Then source .pome-sandbox.env. POME_AUTH_TOKEN is the bearer for EVERY call to the twin and the only one it accepts: any other token, or none, comes back as an opaque 404. Keep it in the shell, and add .pome-sandbox.env to .gitignore. 2. Read the world before you touch it. The routes are Gmail's own, under /gmail/v1/users/me: GET $POME_GMAIL_REST_URL/gmail/v1/users/me/profile GET .../gmail/v1/users/me/messages?q=is:unread GET .../gmail/v1/users/me/messages/msg_support Tell me who is waiting on a reply and what they asked. 3. Answer it, through the twin's own Gmail surface -- the same routes and payloads gmail.googleapis.com takes: messages/send POST with a base64url RFC822 message in `raw` and threadId set to the inquiry's thread, so the reply threads instead of starting a new conversation. Set In-Reply-To and References to the inquiry's Message-ID, and include a From header -- this twin requires one rather than inferring it. modify POST messages/msg_support/modify to remove UNREAD and add the existing Label_follow_up. Send exactly ONE reply. Do not delete anything, and do not touch the other two threads. 4. Read the result back THROUGH THE TWIN, not from your memory of what you sent: GET .../gmail/v1/users/me/threads/thread_support GET .../gmail/v1/users/me/messages?labelIds=SENT 5. Show me the tape, one line per call, in the order they happened: GET $POME_GMAIL_REST_URL/_pome/events 6. Stop the sandbox: npx @pome-sh/cli@latest sandbox stop $POME_SESSION_ID --discard This is a sandbox, not an exam. Nothing here is scored and Pome is not evaluating you. Keep your own evaluator and observability setup exactly as it is -- just read the tape when you are done. ``` **One token opens a sandbox: `POME_AUTH_TOKEN`.** It is the bearer on every call to the twin, REST and MCP alike. On most twins the secrets file also carries a provider-shaped token — `POME_GITHUB_TOKEN=github_pat_…`, `POME_STRIPE_API_KEY=sk_test_…` — and that one is *not* the bearer; it is what the twin serves inside the sandbox. The linear twin ships none at all. Send that provider-shaped token as the bearer, or send no bearer at all, and the proxy answers an opaque `404` reading `No twin pod for this session.` So a 404 on a sandbox you just created is almost always the wrong bearer rather than a dead sandbox. ## The world `sandbox create` boots a Gmail twin from its declared starting state and hands back the URLs that reach it. It also writes the connection secrets to `.pome-sandbox.env` at mode `0600`, and says so on stderr. ```bash theme={"dark"} npx @pome-sh/cli@latest sandbox create --twin gmail \ --secrets-file .pome-sandbox.env --format json ``` ```json theme={"dark"} { "session_id": "ses_PbRt4mOQreRbIA0P", "expires_at": "2026-08-25T10:23:04.482Z", "per_twin": { "gmail": { "api_url": "https://twins.pome.sh/gmail/s/ses_PbRt4mOQreRbIA0P", "mcp_url": "https://twins.pome.sh/gmail/s/ses_PbRt4mOQreRbIA0P/mcp" } }, "provider_credentials": {}, "agent_token": "***redacted***" } ``` That empty `provider_credentials` is not an omission — it is the point. There is no upstream Gmail credential to hand you, so the twin serves its own mailbox and nothing you do here can reach a real one. ```bash theme={"dark"} source .pome-sandbox.env gm() { curl -s -H "Authorization: Bearer $POME_AUTH_TOKEN" \ "$POME_GMAIL_REST_URL$1" "${@:2}" } ``` The world is one mailbox with four threads, two of them unread: ```bash theme={"dark"} gm /gmail/v1/users/me/profile | jq -r '"\(.emailAddress)" + " messages=\(.messagesTotal) threads=\(.threadsTotal)"' gm "/gmail/v1/users/me/messages?q=is:unread" \ | jq -r '.messages[] | "\(.id) \(.threadId)"' ``` ```text theme={"dark"} pome-agent@pome-twin.test messages=5 threads=4 ``` ```text theme={"dark"} msg_support thread_support msg_build thread_build ``` One of those two is a customer waiting on an answer. Read it: ```bash theme={"dark"} gm /gmail/v1/users/me/messages/msg_support | jq -r ' "labels: \(.labelIds | join(", "))", "from: \((.payload.headers[]|select(.name=="From")|.value))", "subject: \((.payload.headers[]|select(.name=="Subject")|.value))", "msg-id: \((.payload.headers[]|select(.name=="Message-ID")|.value))", "body: \(.snippet)"' ``` ```text theme={"dark"} labels: INBOX, UNREAD from: alice@example.com subject: Production export is stuck msg-id: body: Our production export has been stuck for an hour. Can you investigate? ``` Unread, unanswered, and the same every time you create this sandbox — which is what makes anything you observe next reproducible. Now the reply. Gmail takes a whole RFC822 message, base64url-encoded, in `raw`: ```bash theme={"dark"} RAW=$(printf '%s\n' \ 'From: pome-agent@pome-twin.test' \ 'To: alice@example.com' \ 'Subject: Re: Production export is stuck' \ 'In-Reply-To: ' \ 'References: ' \ '' \ 'Looking at the stuck export now - will update within the hour.' \ | base64 | tr -d '\n' | tr '+/' '-_') gm /gmail/v1/users/me/messages/send -X POST \ -H 'content-type: application/json' \ -d "{\"raw\":\"$RAW\",\"threadId\":\"thread_support\"}" \ | jq -c '{id, threadId, labelIds}' gm /gmail/v1/users/me/messages/msg_support/modify -X POST \ -H 'content-type: application/json' \ -d '{"removeLabelIds":["UNREAD"], "addLabelIds":["Label_follow_up"]}' | jq -c '{id, labelIds}' ``` ```text theme={"dark"} {"id":"msg_0000000000000005","threadId":"thread_support","labelIds":["SENT"]} {"id":"msg_support","labelIds":["INBOX","Label_follow_up"]} ``` Read the thread back through the twin's own surface — not out of the response your agent already holds: ```bash theme={"dark"} gm /gmail/v1/users/me/threads/thread_support | jq -r '.messages[] | "\(.id) [\(.labelIds|join(","))] " + "\((.payload.headers[]|select(.name=="From")|.value))"' gm "/gmail/v1/users/me/messages?labelIds=SENT" \ | jq -c '{n: (.messages|length), ids: [.messages[].id]}' ``` ```text theme={"dark"} msg_support [INBOX,Label_follow_up] alice@example.com msg_0000000000000005 [SENT] pome-agent@pome-twin.test ``` ```text theme={"dark"} {"n":2,"ids":["msg_0000000000000005","msg_build_reply"]} ``` **This is the thing worth noticing.** The reply joined Alice's thread rather than starting a fourth one, and `msg_support` lost `UNREAD` and gained a label — read back through two different routes. `SENT` now holds two messages, not one: the mailbox started with a sent reply on the build thread, so **two is the number that means "replied exactly once"**, and that is the number a check has to assert. Stop the sandbox and all of it is gone. ## Read the tape Every call above was recorded by the twin as it happened. This is the part neither a mock nor a test mailbox gives you: an account of the run written by the service, not by the agent. ```bash theme={"dark"} gm /_pome/events | jq -r '.[] | "\(.method|(.+" ")[:5])\(.path|sub("^/s/[^/]+";"")) -> \(.status)"' ``` ```text theme={"dark"} GET /gmail/v1/users/me/profile -> 200 GET /gmail/v1/users/me/messages -> 200 GET /gmail/v1/users/me/messages/msg_support -> 200 POST /gmail/v1/users/me/messages/send -> 200 POST /gmail/v1/users/me/messages/msg_support/modify -> 200 GET /gmail/v1/users/me/threads/thread_support -> 200 GET /gmail/v1/users/me/messages -> 200 ``` Each row carries more than the request line: ```bash theme={"dark"} gm /_pome/events | jq -r '.[] | select(.state_mutation) | "\(.path|sub("^/s/[^/]+/gmail/v1/users/me";"")) \(.status)" + " tool=\(.tool) mut=\(.state_mutation) fid=\(.fidelity)"' ``` ```text theme={"dark"} /messages/send 200 tool=null mut=true fid=semantic /messages/msg_support/modify 200 tool=null mut=true fid=semantic ``` **You should see:** * **Seven rows, in the order they happened.** Read top to bottom and the triage is legible without asking the agent what it did: it looked, it read the inquiry, it replied, it filed, it checked twice. * **`mut=true` on exactly the two writes.** `state_mutation` means the call *landed* — a write the twin refuses reads `false`, not `true`. Send the same message with no `From` header and you get a `400` with `mut=false` beside it. * **`tool=null` on every row.** Unlike the GitHub twin, this one stamps no action vocabulary; calls are identified by method and path. * **`fid=semantic`.** These surfaces carry a full behavioural contract, not a response shape with placeholder values. Now the part that is specific to mail. Ask the recorder what it kept about the message you just sent: ```bash theme={"dark"} gm /_pome/events | jq '[.[] | select(.state_delta) | .state_delta.after.messages[]?] | last | {id, threadId, subject, sentAt, bodyOmitted, body_bytes: .textSha256.size, headerCount}' gm /_pome/events | jq -r '[.[] | select(.state_delta) | .state_delta.after.messages[]?] | last | .textSha256.sha256' ``` ```json theme={"dark"} { "id": "msg_0000000000000005", "threadId": "thread_support", "subject": "Re: Production export is stuck", "sentAt": "2026-07-20T00:00:01.000Z", "bodyOmitted": true, "body_bytes": 63, "headerCount": 5 } ``` ```text theme={"dark"} afcafdd1cb3c9dc55df1511b3d28457b7eb2d4171747d6210a80afac6e7a2260 ``` Two things there are worth a second look. **`bodyOmitted: true`** — the API served you that body a moment ago, but the recorded state keeps a SHA-256 and a byte count in its place, so the evidence a grader reads can prove *which* message was sent without holding what it said. And **`sentAt` is `2026-07-20T00:00:01.000Z`**, one second past the mailbox's own clock rather than the wall clock you ran this at: the twin advances a logical clock per write, so running this twice gives you the same timestamp and the same hash. When you are done, stop the sandbox to free the slot — it also expires on its own 30 minutes after it was created: ```bash theme={"dark"} npx @pome-sh/cli@latest sandbox stop $POME_SESSION_ID --discard ``` ## What you could assert here Nothing on this page was graded, but a tape and a final state are exactly what a graded check reads. The Gmail twin already declares these — no authoring required, and `list_checks` on the Pome MCP prints the full set with what each one actually compares: | Declared check | What it would catch here | | ---------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `gmail.mailbox-label-count` (`pome-agent@pome-twin.test`, `2`, `SENT`) | An agent that replies twice — the retry after a response it thought it lost. The count is exact rather than a lower bound, so three sent messages fail it. And because the mailbox seeds one sent message, an agent that never replies reads `1` and fails too. | | `gmail.message-has-label` (`msg_support`, `Follow Up`) | An agent that answers the customer and never files the mail, so the next person through the inbox cannot tell it was handled. It asks only whether that one label is on the message — it says nothing about the others, so `INBOX` staying put is fine and `UNREAD` staying put would not fail it. The label may be named by its id (`Label_follow_up`) or its display name. | | `gmail.no-unsupported-endpoint` | An agent that guesses at the route prefix. Gmail's paths are easy to get almost right — drop the leading `/gmail` segment and the twin answers `501`, which is exactly what this scans the tape for. | Pointers, not a task. Turning them into a graded exam is the [capstone](/quickstart/claude-code) below, and [Write a task](/docs/authoring-tasks) is where the grammar lives. ## How real is this twin? Every surface you called above is compared against a captured response from the real Gmail API, re-run daily and published. [The Gmail row on status.pome.sh](https://status.pome.sh) carries the current count of twin responses that match, names every divergence we have ruled on by id, and states how old the captured baseline is. Nobody has to take our word for the fidelity, including us — a twin that drifts turns that row red. Which surfaces are covered, and which are shape-only, is on the [Gmail twin reference](/docs/twins/gmail). ## Next: the one graded lesson The same twins, now an exam: a sealed agent under test, a deliberate failing score, and one line of prompt that turns it green. Every route and MCP tool the twin serves, by use case, with its fidelity tier. Turn the checks above into a graded exam for your own agent. Create, list and stop sandboxes — including multi-twin ones.