Show HN: Fastbrowse, Browser agent that picks clicks with a System One model

Browser automation with results you can verify.

Give your agenta browser.

Complete a workflow and get a result backed by quoted page text, with credentials kept out of model prompts. Run locally or in a cloud browser.

Get startedRun it hostedUse it from Codex or Claude Code

MIT licensed.

Interactive replayRecorded task, recreated interface

Task

Sign in as standard_user, add the Sauce Labs Backpack and Bike Light, check out as Ada Lovelace, finish the order and read the total.

Recorded time41.3sModel cost$0.0501Steps19Sign inChooseCartCheckoutVerifyhttps://www.saucedemo.com

Sauce Demo

Usernamestandard_userPassword••••••••••Login

Credentials stay out of prompts

The password is passed by name. --secret password=SAUCE_PASSWORD resolves on the allowed origin only, and the value never enters a prompt.

https://www.saucedemo.com

Products

Sauce Labs Backpack

$29.99

Added

Sauce Labs Bike Light

$9.99

Added

Two items are in the cart.

Choose from real controls

The controls a page offers are indexed into a list. Jev selects the operation and target from that list.

https://www.saucedemo.com

Your cart

Sauce Labs Backpack

$29.99

Qty 1

Sauce Labs Bike Light

$9.99

Qty 1Checkout

Read the page after acting

Both products appear in the cart. The page is observed again before the next action is selected.

https://www.saucedemo.com

Checkout

First nameAdaLast nameLovelacePostcodeEC1A 1BBFinish

Keep control over the order

The caller authorized the order with --authorize. Without authorization, actions classified as irreversible pause for approval.

https://www.saucedemo.com

Order complete

$43.18

The total for the completed order.

"Total: $43.18"Quoted from the checkout overview

An answer you can check

The answer cites the total captured from the checkout overview. Completion also requires evidence that the order was placed.

PlayReplay

Recorded in local Chrome. Cost covers all model calls in this run; there was no cloud-browser charge. This interface is recreated and playback speed is illustrative.

Original recording

The recording runs fastbrowse 0.7.2 in local Chrome against Sauce Demo, a public practice store. The caller authorized the order before the run. The card at the end carries the run's own time, step count and model cost, unrounded.

Replicate this run

The same task and flags, on any machine with uv and a Chrome install. The recording was made on fastbrowse 0.7.2, which is not on PyPI yet, so this installs that build from its commit. Once 0.7.2 is published the long form collapses to uvx fastbrowse==0.7.2.

# Install uv first: https://docs.astral.sh/uv/
export OPENROUTER_API_KEY=...        # or AI_GATEWAY_API_KEY, which also serves the LLM
export SAUCE_PASSWORD=secret_sauce   # public practice credential

uvx --from "git+https://github.com/agent-labs-dev/fastbrowse@6dd7256" fastbrowse \
  "Log in as standard_user, add the Sauce Labs Backpack and the Sauce Labs Bike Light to the cart, and check out as Ada Lovelace, postcode EC1A 1BB. Finish the order and tell me the total." \
  --start https://www.saucedemo.com/ \
  --secret password=SAUCE_PASSWORD \
  --local --authorize \
  --record demo.mp4

--record writes the video, ending on the card with the run's own time, steps and cost. Drop it for a text answer only.

Why fastbrowse.

Compare browser actions, answers and credentials using each project's documentation. See benchmarks for measured speed and cost.

Dimension

fastbrowse

Browser Use agents

Action selection

fastbrowse: Action selectionJev chooses an action and its target from a list of controls on the page.Browser Use agents: Action selectionThe hosted agent generates actions. Ultrafast chooses actions with Jev.

Answers you can check

fastbrowse: Answers you can checkFactual claims in a complete answer cite verbatim quotes from the page. Structured output is schema-validated before your code reads it.Browser Use agents: Answers you can checkThe agent reports the outcome. Browser Use's docs advise checking important actions independently. Ultrafast navigation tests check the final page.

Credentials and control

fastbrowse: Credentials and controlScoped secrets and an authorization gate, built-in Bitwarden, and vault logins through Nebula.Browser Use agents: Credentials and controlProfiles and scoped secret bindings are supported.

Where it runs

fastbrowse: Where it runsThe fastbrowse hosted browser pool, local Chrome, a Browser Use Cloud browser, or a browser you already have open.Browser Use agents: Where it runsAgents and browser infrastructure, including a cloud browser.

Sources: fastbrowse's README comparison, Browser Use's completion guidance and secret bindings.

How fastbrowse works.

fastbrowse lists the page controls, chooses an action, then checks the result.

Choose from indexed controls

Jev chooses the operation and target in one request from a list of controls on the page. Short factual reads can use Jev as well; the LLM plans the task and reads longer passages.

Check the result before finishing

Factual answers cite quotes copied from the page. fastbrowse checks those quotes and the task requirements before reporting the task as complete.

Inspect steps and cost

Each result includes steps, model calls and cost. Compare agents on task success, speed and cost on the benchmarks page.

View benchmarks

Logins and control.

Credentials stay out of prompts

Reuse a saved local or cloud browser profile, or attach to an existing browser. When a credential is needed, fastbrowse resolves the named secret for its permitted website without putting the value in a prompt.

Built-in Bitwarden

Resolve logins from the unlocked Bitwarden CLI by their saved website address, including one-time codes when configured.

Authorization before irreversible actions

fastbrowse asks for approval before actions classified as irreversible. The classifier can miss an irreversible action.

Read the architecture

Run fastbrowse through Nebula.

Nebula uses fastbrowse for browser tasks and shows its steps and results. Connect Bitwarden or 1Password through Nebula to use saved logins on websites you authorize.

nebula.gg

Choose a browser.

Run hosted, run locally, use a cloud browser, or attach to an existing session.

Hosted

Once a live-key run starts on fastbrowse hosted, it executes in a sandboxed Chrome on a managed browser pool, with nothing to install. Run it hosted

Local Chrome

Pass --local to run Chrome on this machine. --headed shows the window, and --profile keeps a signed-in profile on disk.

Cloud browser

The default is a Browser Use Cloud browser, which passes bot checks a fresh local Chrome can fail. It prints a URL to watch the run live.

Attach to a browser

Drive a browser that is already running over --cdp-url or --cdp-port, and leave the browser running after the task. --attach takes over an open window, such as an Electron app.

Cloud extensions

Load up to three ready Browser Use Cloud extensions by ID with cloud_extensions. They are available on a cloud browser only.

MCP and embedding

fastbrowse-mcp serves one browse tool to Claude Code, Claude Desktop, Cursor and other MCP clients. The embedding API exposes browser steps, quotes and model spend.

Use it from your agent.

Install the fastbrowse skill in Codex or Claude Code, then hand your agent a website task. The skill installs instructions only: set the model key and choose a browser as for the CLI. A run is finished only when its status is complete.

Install the skill

npx skills add agent-labs-dev/fastbrowse --skill fastbrowse

Choose your agent when the installer prompts, then start a new agent session.

Ask from Codex

Use $fastbrowse to find the top story on https://news.ycombinator.com/ and return its title with a citation.

Ask from Claude Code

/fastbrowse Find the top story on https://news.ycombinator.com/ and return its title with a citation.

Without a skill

Agents without skill support can run the CLI or connect the MCP server directly. The skill does not authorize writes on its own: say whether the agent may submit a form.

Skill guideSKILL.mdllms.txt

Run a task.

Install the CLI, the Python SDK or the JavaScript SDK, then give fastbrowse a task and a starting page. The JavaScript SDK needs Node.js 20 or newer, and no Python.

CLI usagePython SDKJavaScript SDKEvals & testingcopy

# Install uv first; replace ... with your provider keys
export OPENROUTER_API_KEY=...   # for Jev and the LLM that plans and reads
# export TYPESAFE_API_KEY=...   # optional: use direct TypeSafe for Jev instead
# export AI_GATEWAY_API_KEY=... # Jev backup; also serves the LLM when OPENROUTER_API_KEY is unset
export BROWSER_USE_API_KEY=...  # the cloud browser; or pass --local to use Chrome

uvx fastbrowse "What is the title of the top story right now?" \
  --start https://news.ycombinator.com/

# Without a starting page, fastbrowse proposes the first address from the task
uvx fastbrowse "Find the httpx package on PyPI and report its latest released version."

# Authorization gates and a secret the model only ever sees by name
export SAUCE_PASSWORD=secret_sauce  # public practice credential
uvx fastbrowse "Log in as standard_user and add the backpack to the cart." \
  --start https://www.saucedemo.com/ \
  --secret password=SAUCE_PASSWORD \
  --authorize

# uv tool install fastbrowse  # keep fastbrowse on PATH
# uv add fastbrowse           # or put it in a project

# uv add fastbrowse
# Set the OpenRouter key shown in CLI usage; with no browser given, this runs local Chrome.
import asyncio
from pydantic import BaseModel
from fastbrowse import run_task
from fastbrowse.models import Limits

class Release(BaseModel):
    package: str
    version: str

async def main() -> None:
    result = await run_task(
        "Find the httpx package and report its name and latest released version.",
        start="https://pypi.org/",
        output_schema=Release,
        limits=Limits(max_dollars=0.10),
    )
    if not result.succeeded:
        print(result.status, result.error)
        return
    release = Release.model_validate(result.data)
    print(release.package, release.version)
    for citation in result.citations:
        print(citation.quote, citation.deep_link)

asyncio.run(main())

// npm install fastbrowse zod
// Node.js 20 or newer. No Python and no uv on the machine.
// The npm package installs the native binary for your platform.
// Set OPENROUTER_API_KEY in the environment for Jev and the planning LLM.

import { Fastbrowse } from 'fastbrowse';
import { z } from 'zod';

const Release = z.object({ package: z.string(), version: z.string() });

const fb = await Fastbrowse.start({ local: true });
try {
  const result = await fb.run('Find the httpx package and report its name and latest released version.', {
    start: 'https://pypi.org/',
    output: Release,
    limits: { maxDollars: 0.1 },
    onEvent: event => {
      if (event.type === 'step') console.error(event.step.operation, event.step.target);
    },
    signal: AbortSignal.timeout(120_000),
  });
  console.log(result.status, result.answer);
  if (result.status === 'complete') console.log(result.output.package, result.output.version);
  for (const evidence of result.evidence) console.log(`  "${evidence.quote}" from ${evidence.url}`);
} finally {
  await fb.close();
}

# Working on fastbrowse itself
git clone https://github.com/agent-labs-dev/fastbrowse.git && cd fastbrowse
uv sync --all-extras && cp .env.example .env
# Fill in .env with an OpenRouter key; live runs also need BROWSER_USE_API_KEY.

# Run the fixture evaluation suite
uv run python -m fastbrowse.evals.runner

# Compare task success, time and cost using the published benchmark methods.
# Open Benchmarks on this site for the comparison methods.

# Lint, types, tests
uv run ruff check . && uv run ty check && uv run pytest

Docs and setupGitHub README

Run it hosted.

fastbrowse hosted runs the same agent in managed browsers behind an API. Send a task over HTTPS, follow its steps, and pay per run. Keys, secrets, billing and run history live in the console.

Create an accountAPI docs

Sandboxed browser pool

Once a live-key run starts, it executes in a sandboxed Chrome on a browser pool that fastbrowse runs on Google Kubernetes Engine. There is no browser to install, update or keep alive.

Console and API

Sign in at console.fastbrowse.ai to submit runs, follow their steps and see your balance. Your code sends the same runs to api.fastbrowse.ai.

One request per task

Send a task, and optionally a starting page, to POST /v1/runs at api.fastbrowse.ai. The service runs the browser fleet, model connections, queue and secret store. A test key runs a simulator that returns the same response shapes at no cost.

Pay per run

A run is priced as a per-run fee plus usage, never below $0.005. Runs are paid from a balance that your plan's monthly allowance and any top-ups fill. Top up prepaid credits in the console, and the dashboard shows the balance. An allowance does not roll over; a top-up does not expire.

What a run is charged for

A run is charged when the agent did the work, whatever it proved, including a run that stopped at a sign-in wall or a bot check. A provider outage, an error inside the agent, a cancelled run, a run that expired in the queue and any run made with a test key are not charged.

Plan

Price

Includes

Free

Build an integration and see it work, without a card.

$0

• $2.00 of runs a month
• 1 run at a time
• 7 day history
• Cited answers and step traces

Starter

For a product with browser work in it.

$29 a month

• $35.00 of runs included
• 5 runs at a time
• 30 day history
• Secrets, profiles and schedules
• Webhooks

Pro

For browser work that is part of how the product runs.

$199 a month

• $250.00 of runs included
• 25 runs at a time
• 90 day history
• Session recordings
• Audit log and SSO-ready roles

Scale

Fleet-sized volume with reserved capacity.

$999 a month

• $1,400.00 of runs included
• 100 runs at a time
• 180 day history
• Queue priority
• Shared Slack channel

Enterprise

Dedicated browsers, a contract and a number to call.

By agreement

• Dedicated browser fleet
• Custom retention and residency
• Invoiced billing
• Uptime commitment

Start freeSign in to the consoleFull plan comparisonAPI reference