outcrawl

Agents

Describe a task. An agent does it.

An agent opens a real browser, works through the site and reports back. Watch it live and take over any time. About 15 credits a minute, about $0.07 to $0.08.

Get a free API keyRead the agents docs

1,000 free credits every month. No card needed.

The task, as CLI submits it: outcrawl agent "On books.toscrape.com, open the Travel category and report the cheapest book in it, with its price." \ --budget 16.00 --no-cache --schema "$(cat book.json)"
Submitted202run_54be95df-…id in 992 ms
Run
completed12.6 s5.832706 credits · about $0.0286
  1. 4.0 sgoalopen Travel category (active) · find cheapest book and price (pending)
  2. 4.0 sstep 0navigate https://books.toscrape.com/catalogue/category/books/travel_2/index.html
  3. 8.9 sresult{"price":"£23.21","title":"The Road to Little Dribbling: Adventures of an American in Britain (Notes From a Small Island #2)"}
  4. 8.9 sstep 1extract all books in Travel category with prices; title and price of the cheapest
  5. 11.2 sgoalopen Travel category (done) · find cheapest book and price (done)
  6. 11.2 sstep 2done

Captured from a real run on api.outcrawl.ai on 7 October 2026.

How it works

You ask, it works, it reports back. It runs as long as the task needs.

  1. 01

    Ask in a sentence.

    Say what you want done, the way you would ask a person. Name the fields you want back, and the answer comes back in exactly that shape.

  2. 02

    You get an id straight away.

    A run is a job, not a call you hold open: the id comes back in milliseconds, and the run takes seconds, minutes or hours, as the task needs.

  3. 03

    It works in a real browser.

    It makes a plan, then opens pages, clicks, types and reads, and writes down every step and everything it finds as it goes.

  4. 04

    It reports back.

    You get its answer and a recording of the whole run. A run you stop early keeps everything it already found.

How agents work in the docs

While it runs

Watch it, step in, or leave it to finish. Your agent can do each of these for you.

  • watchUrl

    A link to watch the browser live while it works. The same link plays the recording once it has finished.

    outcrawl_agent

  • status

    Where the run has got to: its plan, what it has spent so far, and its answer once it has one.

    outcrawl_agent_status

  • events

    Every step, finding and change of plan, in order. Pick up from where you last read.

    outcrawl_agent_events

  • results

    What it has collected so far. Records land during the run, so you can read them before it ends.

    outcrawl_agent_results

  • control

    Pause it where it stands, do one step yourself, such as a sign-in or a final confirm, and hand it back. It carries on from the page you left.

    outcrawl_agent_control

  • answer

    When it needs something only you know, it pauses and asks. Answer, and it carries on from where it stopped.

    outcrawl_agent_answer

  • files

    Give it a file to upload, like a resume. The model never reads it, and it is deleted when the run ends.

    outcrawl_agent_file_add

  • cancel

    Stop it at any time. Everything it found stays yours.

    outcrawl_agent_cancel

  • notify

    Have every step, or only the ending, sent to your webhook as it happens. Signed, and retried for hours if your endpoint is down.

    notify.level

Find any earlier run, with what it spent and its recording, with outcrawl_agent_list.

Before it starts

Say where it may go and what it may spend. Everything else has a sensible default.

Set the most it can spendcaps.budget
Default: required
A hard ceiling in credits. When the run reaches it, it stops and keeps what it found. MCP and the CLI fill in 2,000 if you leave it out.
Start signed inprofile
Default: none
A saved profile's logins and cookies, so it starts where you left off.
Browse from a countryexit
Default: none
A country, region or city it browses from.
Rules it must followrules
Default: workspace rules
Plain sentences like “Never make a payment.” Each comes back marked enforced, a gate that blocks the action, or guidance. Your agent sets them with outcrawl_rules_set.
Answer in named fieldsschema
Default: none
A JSON Schema. The answer comes back in that shape, and filling it is what counts as done.
Type in a saved password or cardsecrets
Default: none
Credentials you stored from the CLI or an SDK, typed into the page without the model ever seeing them. Only the ones the run names.
Reach past the browserintegrations
Default: none
Tools from an MCP server you connected, such as Slack or Gmail. Only the ones the run names.
Skip the recordingrecord
Default: true
Every run is recorded unless you send false. It costs the same either way.

Where agents stop

  • It takes no new instructions once it starts. Take control to step in, or cancel and start again.

  • A question waits for you, not forever. If nobody answers in time, it carries on without the answer.

  • Saved credentials and connected tools are set up by a person, never by an agent over MCP.

  • To read a page you can name, Scrape is far cheaper.
Every option in the docs

What you can build

Hand off the jobs that need a person. Real examples you can hand to your agent.

Docs

A monitor watches a booking page; when a slot opens, your rules pick it and an agent books it, or holds it for you to confirm in the live view.

example
Watch the Riverside Clinic dermatology page every 10 minutes with my clinic profile, and tell me on Slack when a slot opens. outcrawl_monitor_create: checking every 10 minutes, 1 credit a check. The Tuesday 08:40 with Dr Osei fits. Get it to the last screen and let me confirm it myself. outcrawl_agent: submitted · watch it live. outcrawl_agent_control: held on the form, with the live view link. Done, I confirmed it. outcrawl_agent_cancel: run ended, nothing more billed.
80 blocked sites, run 7 October 2026

Benchmark

It works on the sites where other agents get stopped.

Browser Use built this test from the 80 sites where AI agents got blocked most often in their own traffic. We ran it on Outcrawl and scored it with their own code. The next best result was 81.4%.

Pages reached

92.5%

Mean of 3 runs on 71 scored tasks, 95% interval 91.5% to 94.4%.

  1. Outcrawl92.5%
  2. Browser Use Cloud81.4%
  3. Anchor76.8%
  4. Kernel67.1%
  5. Browserless54.0%
  6. Steel47.2%
  7. Browserbase41.9%
  8. Hyperbrowser39.9%

Mean pass rate, 71 scored tasks, Browser Use’s own scoring code. Ours: 3 runs, 7 October 2026.Method

Pricing

Billed for what the run uses. You set the most it can spend.

Tool

Dollars on

  • Agents

    A task in plain English, done in a real browser.

    About 15 credits a minute

    about $0.07 a minute

  • Browsers

    A real browser in the cloud that you control.

    1 credit a minute

    about $0.29 an hour

  • Proxy usage

    Data through a proxy, on browser sessions and agent runs.

    2,000 credits a GB

    $9.80 a GB

Agent runs are billed for the AI model's usage, browser time and proxy usage, in credits. The estimate is a typical run; the actual cost depends on the task. The middle half of runs land between 10 and 25 credits a minute on Builder. Model usage converts at your plan's price per credit, so on Pro and Scale the same minute is more credits.

A run of 5 minutes, estimated

Builderabout 15 credits a minute
75 creditsabout $0.37
Proabout 20 credits a minute
100 creditsabout $0.40
Scaleabout 25 credits a minute
125 creditsabout $0.40

About the same in dollars on every plan, because cheaper credits mean more of them. An estimate: the real cost depends on the task.

  • The budget is a hard ceiling: a run with a budget of 100 credits never bills more, $0.49 on Builder.

  • While you have control, only browser time and proxy usage are billed, no model usage.

  • Cancelling stops the spending. What the run already did stays billed.

See full pricing

FAQ

Questions about agents