How is a run charged?

A per-run fee plus usage, never below $0.005. A run is charged for the work the agent did, whatever the run proved — including runs that ended unverified, were blocked, or stopped at needs_confirmation or needs_login. Platform errors, unavailable providers, cancelled runs, runs expired in the queue and test-key runs are not charged. Monthly allowances do not roll over; top-ups do not expire.

Do I need a card to start?

No. The Free plan needs no card and carries $2.00 of runs a month (at most 400 runs, since none costs less than $0.005), one run at a time, a 7-day history, and a cap of 500 runs a month including top-ups.

How do I use it from my agent or code?

Install the fastbrowse skill in Codex or Claude Code (npx skills add agent-labs-dev/fastbrowse --skill fastbrowse), connect fastbrowse-mcp to Claude Code, Claude Desktop, Cursor or other MCP clients, or use the CLI, Python SDK or JavaScript SDK (Node.js 20+, no Python needed). An embedding API also exposes browser steps, quotes and model spend.

fastbrowse — Browser automation with results you can verify

fastbrowse is an MIT-licensed browser automation tool built around verifiable results: it completes a workflow in a real browser and returns an answer backed by verbatim quotes from the page, or schema-validated data when asked. Its action model, Jev, selects each operation and target from a list of indexed page controls, and an LLM plans the task. Credentials are resolved by name for their permitted origin only and never enter model prompts, with built-in Bitwarden support and 1Password via Nebula. Runs can execute on the fastbrowse hosted pool, in local Chrome, in a Browser Use Cloud browser, or against an existing browser over CDP, and every result reports steps, model calls and cost. It ships as a CLI, Python and JavaScript SDKs, an MCP server, and a skill for Codex or Claude Code, with a hosted, usage-priced service on top.

Key Features

Answers backed by verbatim page quotes

Factual claims in a completed answer cite verbatim quotes copied from the page, and fastbrowse checks those quotes and the task requirements before reporting the task as complete.

Schema-validated structured output

Pass a Pydantic model (Python) or Zod schema (JavaScript) and get data that is validated before your code reads it, alongside citations with deep links.

Action selection from indexed page controls

The controls a page offers are indexed into a list, and Jev chooses the operation and target in one request from that list; an LLM plans the task and handles longer reads.

Credentials kept out of model prompts

Secrets are passed by name and resolved only on their permitted origin, so values never enter a prompt. Built-in Bitwarden CLI resolution includes one-time codes; 1Password works through Nebula.

Authorization gate for irreversible actions

Actions classified as irreversible pause for approval unless the caller passes --authorize; the vendor notes the classifier can miss an irreversible action.

Four ways to run a browser

fastbrowse hosted pool, local Chrome (--local, --headed, --profile), a Browser Use Cloud browser (the cloud default, which passes bot checks a fresh local Chrome can fail and prints a URL to watch the run live), or attach to a running browser over --cdp-url/--cdp-port/--attach, including Electron ap

Step, cost and recording transparency

Each result includes steps, model calls and cost; --record writes a video of the run ending on a card with the run's own time, step count and unrounded model cost.

Agent and SDK integrations

CLI, Python SDK, and JavaScript SDK (Node.js 20+, native binary, no Python needed); fastbrowse-mcp serves one browse tool to Claude Code, Claude Desktop, Cursor and other MCP clients; a skill installs into Codex or Claude Code.

How It Works

Choose from indexed controls

The page's controls are indexed into a list, and Jev picks the operation and target in one request; short factual reads can also use Jev, while the LLM plans the task and reads longer passages.

Check the result before finishing

Factual answers cite quotes copied from the page, and fastbrowse verifies those quotes and the task requirements before reporting completion; the page is re-read after each action.

Inspect steps and cost

Each result includes steps, model calls and cost, and runs can be recorded to video for replay.

Pros & Cons

Pros

Results are verifiable: answers cite verbatim page quotes that are checked before completion, and structured output is schema-validated
Credentials never enter model prompts; secrets are scoped to an allowed origin, with Bitwarden support built in and 1Password via Nebula
MIT-licensed and self-hostable, with a managed hosted option that runs the browsers, model connections, queue and secret store
Runs in local Chrome, the hosted pool, a Browser Use Cloud browser, or an existing browser/Electron app over CDP
Per-run cost, step count and duration are reported, and runs can be recorded to video or replayed

Who It's For

Best for

Developers wiring browser tasks into agents via the CLI, Python/JavaScript SDKs, MCP servers, or the Codex/Claude Code skill
Teams that need answers with citations or schema-validated data rather than a plain agent self-report
Login-heavy workflows where credentials must stay out of model prompts
Automation where you want to inspect every run's steps, model calls and cost, and replay it on video

Not ideal for

Automation you want with no LLM usage cost — every run carries a per-run fee plus usage
High-volume unattended workloads on the Free plan (1 concurrent run, 500-run monthly cap)

Use Cases

Automating login-protected web workflows with cited results

Extracting schema-validated structured data from websites

Giving coding agents a browser via MCP or the Codex/Claude Code skill

Auditable, recorded runs with per-run cost and step traces

Resolving logins from Bitwarden or 1Password without exposing secrets to prompts

Pricing

fastbrowse hosted is pay-per-run: a per-run fee plus usage, with no run priced below $0.005. Each plan carries a monthly run allowance that does not roll over (purchased top-ups do not expire), and higher plans raise concurrency, request rate and run-history retention. Runs are charged for the work the agent did even if the run ended unverified, was blocked, or paused at needs_confirmation/needs_l

Free

$0/month

No card required. $2.00 of runs a month, 1 run at a time, 60 requests/min, 500-run monthly cap, 7-day history, cited answers and step traces.

Starter

$29/month

$35.00 of runs included, 5 runs at a time, 300 requests/min, 30-day history, secrets, profiles and schedules, webhooks.

Pro

$199/month

$250.00 of runs included, 25 runs at a time, 1,200 requests/min, 90-day history, session recordings, audit log and SSO-ready roles.

Scale

$999/month

$1,400.00 of runs included, 100 runs at a time, 6,000 requests/min, 180-day history, queue priority, shared Slack channel.

Enterprise

By agreement

Dedicated browser fleet, custom retention and residency, invoiced billing, uptime commitment.

FAQ

What does fastbrowse do?

It gives your agent a browser: it completes a workflow and returns an answer backed by quoted page text, or schema-validated data on request, with credentials kept out of model prompts. The hosted version runs the browsers, model connections, queue and secret store, and the open-source project can also be run yourself.

How are results verifiable?

Factual claims in a completed answer cite verbatim quotes copied from the page, and fastbrowse checks those quotes and the task requirements before reporting completion. Structured output is schema-validated (Pydantic in Python, Zod in JavaScript) before your code reads it.

How are logins and credentials handled?

Secrets are passed by name and resolved for their permitted website only, so values never enter a prompt. fastbrowse can resolve logins from an unlocked Bitwarden CLI, including one-time codes, and connect 1Password through Nebula. Actions classified as irreversible pause for approval unless the caller passes --authorize.

Is fastbrowse open source?

fastbrowse