Describe the job. Scribe drives.

Scribe is Writ’s own AI. You say what you want in plain words; it opens a real browser, finds the page, validates a selector on it, and builds the monitor, the workflow or the crawl. It works in turns, and it pauses before anything that costs you money or touches your accounts.

approval ▸ it pauses

It stops and asks. That is the design.

A mission runs in turns. When it needs something from you, it stops: the session goes to awaiting_input and carries a pending_request describing what it is waiting for. There are three shapes of pause:

capabilities ▸ one tool a turn

What you can ask it to do.

Scribe picks exactly one tool per turn, from a fixed set. It cannot invent a tool, so the list below is the whole of what a mission can do:

venues ▸ two, not one

Cloud Scribe and desktop Scribe are not the same.

There are two implementations, and they differ in what they are allowed to build. Read the row that matches where you are running before you plan a mission around it.

On the desktop agent

The same shape, on loopback, against the agent running on your own machine — and free, because the browser is yours:

session ▸ what you poll

The mission runs in the background. You poll it.

Start answers straight away with a session id and a poll URL — a browser mission outlives an HTTP request, so a handle is the honest answer. Every poll returns the same projection:

Start, poll, answer

A mission end to end, over plain HTTP.

# Describe the job. goal is 3-2000 characters; url is an optional seed.
curl -X POST https://api.usewrit.app/api/ai-concierge/start \
  -H "Authorization: Bearer $WRIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "goal": "Watch the price of the 15-inch model and alert me under $1,200",
    "url": "https://example.com/laptops/15"
  }'

{
  "session_id": 4021,
  "status": "planning",
  "poll_url": "/api/ai-concierge/4021"
}

// The mission runs in the background from here. Poll poll_url.
// Every poll answers the same projection:
{
  "session_id": 4021,
  "status": "awaiting_input",
  "phase": "proposing",
  "goal": "...",
  "platform": "cloud",
  "progress_message": "Found the price on the page.",
  "transcript": [],
  "thoughts": [ { "tool": "propose_selectors", "thought": "The price sits in the buy box.", "ts": "..." } ],
  "plan": {},
  "pending_request": { "kind": "text", "requests": [] },
  "resources": {},
  "turn_seq": 3,
  "tokens": { "input": 8120, "output": 940, "credits": 29 },
  "error_message": null
}

# Poll until status is awaiting_input, then answer the CURRENT turn.
curl https://api.usewrit.app/api/ai-concierge/4021 -H "Authorization: Bearer $WRIT_API_KEY"

# turn_seq is an optimistic lock: send back the one you just read.
# A stale turn_seq answers 409 — re-read and answer the current turn.
curl -X POST https://api.usewrit.app/api/ai-concierge/4021/respond \
  -H "Authorization: Bearer $WRIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"turn_seq": 3, "answers": {"threshold": "1200"}}'

# Ask it a follow-up about what it just built.
curl -X POST https://api.usewrit.app/api/ai-concierge/4021/ask \
  -H "Authorization: Bearer $WRIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "How do I call this endpoint from my own code?"}'

# Attach a saved login identity. null clears it.
curl -X POST https://api.usewrit.app/api/ai-concierge/4021/persona \
  -H "Authorization: Bearer $WRIT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"persona_id": 12}'

# Stop mid-turn (the mission parks, resumable) — or end it.
curl -X POST https://api.usewrit.app/api/ai-concierge/4021/interrupt -H "Authorization: Bearer $WRIT_API_KEY"
curl -X POST https://api.usewrit.app/api/ai-concierge/4021/cancel    -H "Authorization: Bearer $WRIT_API_KEY"

# Every mission you have run.
curl https://api.usewrit.app/api/ai-concierge -H "Authorization: Bearer $WRIT_API_KEY"

respond ▸ answering a pause

Answering the turn in front of you.

A response carries turn_seq and an answers object keyed by the fields the pause asked for. turn_seq is an optimistic lock: send back the one you read. If the mission has moved on, you get 409 — refresh and answer the current turn instead of overwriting a newer one.

The same, on your own machine

Loopback, a local token, and the desktop’s narrower scope — watch and notify.

curl -X POST http://127.0.0.1:8131/v1/ai-concierge/start \
  -H "Authorization: Bearer $WRIT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"goal": "Tell me when this page changes", "url": "https://example.com/status"}'

curl http://127.0.0.1:8131/v1/ai-concierge     -H "Authorization: Bearer $WRIT_TOKEN"
curl http://127.0.0.1:8131/v1/ai-concierge/7   -H "Authorization: Bearer $WRIT_TOKEN"

curl -X POST http://127.0.0.1:8131/v1/ai-concierge/7/respond \
  -H "Authorization: Bearer $WRIT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"turn_seq": 2, "answers": {"confirm": true}}'

curl -X POST http://127.0.0.1:8131/v1/ai-concierge/7/interrupt -H "Authorization: Bearer $WRIT_TOKEN"
curl -X POST http://127.0.0.1:8131/v1/ai-concierge/7/cancel    -H "Authorization: Bearer $WRIT_TOKEN"

limits ▸ plainly

What Scribe will not do.

Worth reading before you plan around it. None of these are settings you can change:

1. It does not place the order — For a purchase, it drafts the checkout up to the pay button and stops there. The last click is yours.
2. It never asks for a card number — Payment is collected as a choice — a card already saved on the merchant, your browser’s own autofill on your own machine, or a virtual card. The model cannot ask for a number. If you take the advanced path and enter one yourself, it is sealed into the vault as it arrives and only a placeholder reaches the plan.
3. It cannot invent a tool — One tool per turn, from the fixed set on this page. There is no free-form action.
4. It does not run self-hosted — Scribe is not part of the self-hosted coordinator.
5. Desktop Scribe is narrower — No approved-action drafting, and no multi-site batch monitors.
6. It does not stream its arguments or results — The visible reasoning is the tool it chose and a short thought. Nothing more.

plan ▸ what it costs

What it needs, and what it spends.

Scribe needs a plan that includes AI assist. Token spend is metered and shown inside the session itself — the tokens field carries input, output and the credits consumed so far, so a long mission is never a surprise at the end of it. A mission on your own machine drives your own browser, so the browser time is free; the AI tokens are still metered.

reference ▸ next

What Scribe builds for you

faq

Scribe questions, answered.

end ▸ ask it

Describe one job and watch it pause.

The fastest way to understand the approval model is to start a mission and read the first pending_request it stops on.