Browser automation tool fastbrowse leaves credentials out of model prompts and cites page text in its answers

fastbrowse is a new browser automation tool that gives agents a browser to complete workflows and return results backed by quoted page text. It keeps credentials out of model prompts, runs locally or in a cloud browser, and records interactive replays of every task.

The tool completed a demo task - logging into a retail site, adding two products, checking out, and reading the order total - in 41.3 seconds across 19 steps. The model cost for that run was $0.0501. The answer cited the total directly from the checkout page: "Total: $43.18".

How credential handling works

Passwords are passed by name rather than value. A flag like --secret password=SAUCE_PASSWORD resolves only on the allowed origin, and the actual password never enters a model prompt. For teams using password managers, fastbrowse includes built-in Bitwarden support that resolves logins by their saved website address, including one-time codes when configured.

Before any action classified as irreversible, fastbrowse pauses for approval. The caller authorizes the action with --authorize. The documentation notes the classifier can miss an irreversible action, so the gate is a layer rather than a guarantee.

How it selects and verifies actions

Instead of generating freeform browser commands, fastbrowse indexes every control on a page into a list. Its model, called Jev, chooses an operation and target from that list in a single request. After each action, the page is observed again before the next selection is made. This loop continues until the task is complete.

Factual answers cite verbatim quotes copied from the page. fastbrowse checks those quotes and the task requirements before reporting completion. For structured output, responses are schema-validated before your code reads them.

Running it yourself

A single CLI command replicates the demo run on any machine with uv and Chrome installed. The tool installs from a specific GitHub commit since the current release is not yet on PyPI. The same task runs with identical flags.

fastbrowse also ships as a Python SDK and a JavaScript SDK. The JavaScript package needs Node.js 20 or newer and no Python at all - it installs a native binary for the platform. An MCP server exposes one browse tool to Claude Code, Claude Desktop, Cursor, and other MCP clients. A skill for Codex and Claude Code installs instructions only, with no automatic write authorization.

Browser options include a hosted sandboxed Chrome pool on Google Kubernetes Engine, local Chrome with --local, a Browser Use Cloud browser that passes bot checks, or attaching to an already-open browser over CDP. Cloud extensions load by ID on cloud browsers.

Pricing and plans

Hosted runs are priced as a per-run fee plus usage, never below $0.005. A free plan includes $2.00 of runs per month with one run at a time and seven-day history. Paid tiers start at $29 per month for $35.00 of included runs, scaling to enterprise plans with dedicated browser fleets and invoiced billing. Runs are charged only when the agent did the work - provider outages, cancelled runs, and test-key runs incur no charge.

Why this matters for IT and development teams

Browser automation tools often leak credentials into model contexts or produce answers that cannot be traced back to page content. fastbrowse addresses both problems with named secrets and quoted citations. For developers building agents that need to interact with web applications - filling forms, extracting data, completing multi-step workflows - the combination of verifiable answers and credential isolation reduces the risk of silent failures. Teams can inspect every step, see exact model costs, and replay recordings to debug what happened. The SDKs and MCP server mean the tool slots into existing agent pipelines without replacing the stack.