Browser agents prepare real work on the web through one API. Every end user has a persistent browser and Memory of their own, a person can take over whenever needed, and nothing is submitted until it is approved.
Find one-way flights from Mexico City to New York on 20 November 2026 for one adult in economy class. Stop when the flight options are visible. Do not book anything.
0.0 s
The same application form, the same data, filled by Machinus Agent and by Astra with Codex computer use. A 4-step form with 22 fields and 15 kinds of control: text, long text, date, dropdowns, type-ahead, counter, slider, chips, checkboxes and switches.
4.3xfaster
19.4 s against 83.1 s for the same form.
74xcheaper
$0.03 against $2.22 a form.
100forms for $3
The same $3 buys 1 form from Astra.
Machinus Agent
Astra, Codex computer use
The same form and the same data, one run each. The time is from the first action to the finished form. Astra's $2.22: $0.23 is the task itself; $1.81 is its cached context. Against Astra's cost for the task alone, without its cached context, Machinus is 7.7x cheaper.
BUILT FOR SPEED
Machinus Fast runs an open-weight model served by a managed inference provider. It reads the screen and picks the next action, with form navigation and the user's Memory already fetched before the browser opens.
step speed
Memory and navigation are ready before the first step.
PAY FOR WHAT RUNS
A small, fast model and a persistent browser keep each run light. Usage is recorded as browser minutes, model tokens and Memory searches, so cost follows what was actually used.
Usage ledger
Charged from the ledger, by what was used.
INTELLIGENT EXECUTION
Persistent, approval-gated execution preserves the browser state when a person needs to decide, sign in, or provide a private detail.
BROWSER IDENTITY
Each account gets a stable, honest browser persona: a real display, software WebGL, a consistent fingerprint, and the timezone and language of its proxy country. Nothing pretends to be a different operating system.
Browser persona
looks like a real browser
KEEPS RUNS MOVING
reCAPTCHA v2 and v3 and Cloudflare Turnstile are solved through CapSolver, and every solve is counted in the run's usage.
solvingsolved
Solved through CapSolver
Every solve is counted in the run's usage.
READ-ONLY SOURCES
When Memory has nothing, the agent can read from the apps an end user connected, such as email or documents. Reads only, scoped to that account, and shown as sources in the run.
Sources used in this run
Memory had nothing, so the agent looked it up.
ACCOUNT-SCOPED
Each account gets its own Memory and its own persistent browser profile. Facts, documents and sign-ins stay with that user between runs.
One end user's workspace
ISOLATED CONCURRENCY
Run account-isolated browser work in parallel while keeping each customer’s context, approval, and audit trail intact.
RELIABLE DELIVERY
Durable runs, checkpoint-aware recovery, and explicit status events keep your product informed when execution changes state.
RUN
durable
CHECKPOINT
saved
EVENT
delivered
BRING YOUR OWN ACCOUNT
Connect AI subscriptions your organization owns, then lease healthy accounts per run instead of funding a separate model-credit wallet.
Account holder
Personal provider connection
An account holder can use their own subscription.
Organization pool
Lease healthy capacity without another model-credit wallet.
API SURFACES
Browser Agent API
POST /api/fast/v1/runs
{
"url": "https://forms.example.com/apply",
"goal": "Prepare the application. Do not submit.",
"account": "user_123"
}response shape
status needs_input | ready
questions what is missing, if anything
preview_digest sha256:…
A key can only call the surfaces its scopes allow.
01
POST /api/fast/v1/runs
Hand the agent a form and a goal. It fills what Memory and your context support, asks for what is missing, and stops ready for review. Nothing is submitted until you approve.
02
PUT /api/fast/v1/accounts/{id}
One account for each of your own users, by your id, so no user ever sees another's data.
03
PUT /api/fast/v1/accounts/{id}/memory
Save facts and documents for an end user, then search them. Every run starts from what is already known.
04
POST /api/fast/v1/search
Fresh public-web results without opening a browser, ready to ground an answer.
05
POST /api/fast/v1/fetch
Read pages you already know the address of, and tell when one changed. Your scheduler decides when to check.
06
POST /api/fast/v1/automations
Save a run once and make it again whenever you like, with a note of whether the result changed.
07
POST /api/fast/v1/webhooks
Be told when a run needs an answer, is ready, or finishes, instead of polling.
PLATFORM
Infrastructure access
$99per month
Charged monthly. No annual minimum.
Transparent dollars, provider-metered costs, and no expiring credit balance.
An API that does work in a real cloud browser. Give it a goal and it carries it out and returns the result. It also has Search and Fetch for the public web, and Memory for each of your users' facts and documents.
Simple tasks, like grabbing data from a page or running a search and reading the results. Mid-complexity work, like long applications and forms with document uploads such as a resume or pitch deck, booking a trip end to end (a person completes the payment through a takeover link), comparing flights and prices, or requesting a quote. We expect it to handle far more complex flows too, and we are still testing those.
Not by default. A run fills everything in and stops in front of the final button until a person approves the exact values. To let runs finish on their own, set the run's mode to complete: it submits only when its safeguards pass, and otherwise waits for you.
Create a scoped API key and call the REST API, using the OpenAPI spec as the contract. Runs report through signed webhooks or an event feed. The docs include a prompt for your AI coding agent, so a first run takes minutes.
In credits, for what you use. A run is charged once when it ends, with no flat fee per run, and Search, Fetch and Memory calls are charged per call. Your dashboard shows the credits each endpoint used.
From your goal and that user's Memory: saved facts, uploaded documents and, if you enable it, read-only access to their connected apps. It never invents a value. If something required is missing, the run stops and returns the questions to ask.
Each end user gets their own Memory and their own persistent browser profile, so cookies and sign-ins never mix. You can require recorded consent first, and deleting a user erases their Memory and profile.
The run pauses and asks for a takeover. Send your user a link, they finish that step in the live browser without needing an account, and the run continues.
No. Machinus reads the page's structure (controls, labels, roles and state, including frames and shadow DOM) and acts on those controls directly, which makes it fast and precise. Pages whose content exists only as pixels, such as text in an image, a canvas or a map, give it nothing to read, and unusual widgets can vary by site, so try your own flows first. If it cannot confirm the page shows what you asked for, the run asks for input instead of claiming success. We are in beta and widening what it handles all the time: if a page gives you trouble, send it to us and we will have it fixed within 10 minutes.