Overview
Browser Use is an AI browser agent from Browser Use, Inc.: you describe a task, and a model completes it in a real browser. Per the README, the model writes Python in one persistent REPL, using the browser-harness helpers and the Chrome DevTools Protocol, and the library runs on Pydantic AI since version 1.0.
The terms of service say Browser Use is headquartered in the United States. The repository was created on 31 October 2024 and was last pushed to on 19 November 2026; the latest release, 1.0.6, is from 16 November 2026, so development is active. The README tells users coming from the 0.x versions to read the migration guide.
The company sells three things: the open-source library, Browser Use Cloud (stealth browsers with CAPTCHA solving, plus a hosted agent API) and enterprise deployments. Only the library is open source.
How to Get It
Browser Use has no installer: you use it through a hosted API, a CLI or a Python library, and the library needs Python 3.11 or newer.
Route | How to get it |
|---|---|
| Hosted cloud | Browser Use Cloud: API, $1 free credit |
| CLI | uv tool install browser-use==1.0.6 |
| Python library | uv add 'browser-use==1.0.6' or pip (Python 3.11+) |
| Website | browser-use.com |
| Source code | GitHub (MIT) |
| Privacy policy | browser-use.com/privacy |
Version reviewed: 1.0.6, released on 16 November 2026 per GitHub Releases and matching pyproject.toml; last upstream update (push) on 19 November 2026. The README advises pinning an exact version, because releases move fast and can break things.
# CLI (uv)
uv tool install browser-use==1.0.6
browser-use skill install
# Python library (Python 3.11+)
uv add 'browser-use==1.0.6'
export BROWSER_USE_API_KEY=bu_...Features
Everything below is a maker statement from the README or website; none of it was tested by PromptQuorum.
- An agent that writes Python. The model works in a REPL against the browser through raw Chrome DevTools Protocol calls and ends with a result.
- Several browser modes. A Browser Use Cloud browser (the default), a fresh local Chrome or Chromium with a temporary profile, your own open Chrome with your logins, or a copy of one of your Chrome profiles.
- Choice of model. The hosted bu-4.0 family by default, built-in OpenAI, Anthropic and Google classes, any Pydantic AI model string, or an OpenAI-compatible server such as Ollama.
- Typed results, uploads and downloads. Structured output through a schema, file uploads to pages and downloaded files collected in a workspace folder.
- Secrets and limits. Domain-scoped secrets masked in logs and results, plus allowed and prohibited domain lists. The README states that neither is a sandbox.
- Custom tools and web search. Python functions can be registered as agent tools; an optional web search is billed per search to your Browser Use key.
- A judge step. After a run, one more request to the same model rates the outcome against the task, using the trace and, unless disabled, screenshots.
- CAPTCHAs. Cloud browsers solve CAPTCHAs automatically; the README adds that results depend on the site and that no configuration guarantees success.
Privacy and Data Flow
Browser Use sends what the agent sees to a model, so page content, screenshots and your task text leave your computer unless both the browser and the model are local. In the default setup the model is the maker's hosted gateway and the browser runs in Browser Use Cloud; with a local Chromium and an Ollama model, both stay on your machine.
The privacy policy names Browser Use, Inc. as the responsible entity and lists user-provided inputs and prompts among the data it collects; card payments go through Stripe. The README adds that anonymous telemetry events are recorded when telemetry is enabled, and that the judge step can send screenshots to the model.
Two cautions come from the README: the live-view link of a cloud browser lets anyone who has it control that browser, and driving your own open Chrome gives the agent access to the sites you are logged into. PromptQuorum has not inspected network traffic.
📌Note: These are the maker's statements, not audit results. For work or customer accounts, read the privacy policy and terms first and use a separate browser profile. Automating a website can breach its terms of service, so check them before running an agent against a site, especially with CAPTCHA-solving or stealth features.
Pricing and License
The library is free; the cloud browser and hosted agent are paid by usage after a small credit.
The website lists $0.06 per browser hour for stealth browsers and $0.006 per task (median) for web agents, and says new accounts start with $1 free, one per person, with no expiry. The README names the same $1 credit for new Google, GitHub or Microsoft sign-ups. Enterprise terms such as SOC 2, HIPAA with a BAA, EU data residency and on-premise deployment are offered through sales. Model usage through your own provider is billed by that provider.
Option | Price | Notes |
|---|---|---|
| Library | Free | MIT license, runs on your machine |
| Cloud browsers | $0.06/browser hour | Stealth, CAPTCHA solving |
| Web agents | $0.006/task (median) | Hosted agent API |
Prices are the maker's figures in US dollars, read on 19 November 2026. The task price is described as a median, so individual tasks can cost more or less.
Who Should Use It
- Python developers who want an AI model to automate web tasks and accept a young, fast-moving library.
- People who already run Claude Code, Codex or a similar agent and want to give it browser access through the CLI, as the README describes.
- Teams that want to try a local Ollama model driving a local browser, accepting that quality with small local models was not tested here.
Limits and What We Could Not Verify
Browser Use is a poor fit if page content must not reach a third-party model, and many details remain unverified.
- Not a fit: tasks on accounts where an AI acting for you is not allowed, or where a wrong click costs money or data; the README itself says its safeguards are not a sandbox.
- Not a fit: users who want a double-click installer; the sources show none.
- Unverified: task success rates, speed and cost per task. The README shows a benchmark chart that PromptQuorum did not reproduce.
- Unverified: how well local models drive the browser. The README shows the Ollama setup but gives no quality figures.
- Unverified: stealth, CAPTCHA solving and bot-detection claims, which depend on the site and are the maker's statements.
- Unverified: enterprise security statements and data-residency options, which are offered through sales; no audit report was seen.
Alternatives
Browser Use is a browser-specific agent; the alternatives in the same directory segment are broader computer or coding agents. PromptQuorum has not compared them hands-on.
Browser Use
- Noted in sources:
- Browser-only agent library and CLI; MIT; optional paid cloud
Articles about Browser Use (1)
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
Also mentioned in:
- Cline Review: The Open-Source Agent That Grew From Claude DevUpdated September 5, 2026
- OpenHands Review: The Open-Source Answer to DevinUpdated September 5, 2026
Local Operator
- Noted in sources:
- Desktop app for agents that run code on your computer; MIT
Articles about Local Operator (2)
- Local Operator Review: Free Desktop AI Agent for Your ComputerUpdated October 24, 2026
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
Open Interpreter
- Noted in sources:
- Lets a model control your computer and run code
Articles about Open Interpreter (3)
- Open Interpreter Review: Same Name, Rust Core, New LicenseUpdated September 3, 2026
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
- Nanobot Review: A Self-Hosted AI Agent FrameworkUpdated October 30, 2026
Also mentioned in:
- Phone Harness Review: Let a Coding Agent Use Your iPhone or Android PhoneUpdated November 8, 2026
- Goose Review: Block's Open-Source AI Agent, Now Under the Linux FoundationUpdated October 30, 2026
- Kruvo Review: A Desktop AI Coworker That Asks Before It ActsUpdated October 24, 2026
- Multica Review: A Workspace for AI Coding AgentsUpdated September 14, 2026
- Atomic Agent Review: Local AI Agent CLI, Free & MIT-LicensedUpdated September 12, 2026
- PhoneClaw Review: The On-Device iPhone AI AgentUpdated September 12, 2026
Agent Zero
- Noted in sources:
- Self-hosted agent in Docker with a web interface
Articles about Agent Zero (2)
- Agent Zero Review: Features, Install, AlternativesUpdated September 11, 2026
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
Also mentioned in:
- TecAdRiseBot Review: Free Windows Agent Desktop on Cursor Cloud ModelsUpdated October 28, 2026
- Coil Review: A Mac Agent Harness That Shows Which Model AnsweredUpdated October 24, 2026
- Kruvo Review: A Desktop AI Coworker That Asks Before It ActsUpdated October 24, 2026
- IntentKit Review: Self-Hosted Agent Cluster With Built-In APIsUpdated September 19, 2026
- Local Deep Research Review: Self-Hosted, Cited AI Research AgentUpdated September 19, 2026
- PhoneClaw Review: The On-Device iPhone AI AgentUpdated September 12, 2026
FAQ
Is Browser Use free?
The Python library is free under the MIT license. Browser Use Cloud and the hosted models are billed by usage, with $1 of free credit for new accounts per the website, and API providers charge separately if you use your own keys.
Can Browser Use run with a local model?
Yes, per the README FAQ: use a local Chromium and an OpenAI-compatible local endpoint such as Ollama on localhost port 11434. PromptQuorum has not tried it, and the sources give no minimum model size.
Does it use my own Chrome and logins?
It can: the Chrome mode drives the Chrome you have open, with your logins, after you enable remote debugging once in chrome://inspect. A temporary copy of a profile is also possible. This gives the agent access to those sessions, so use a separate profile for risky tasks.
Which version should I use?
The reviewed release is 1.0.6 from 16 November 2026. The README says to pin an exact version rather than a range, and the 1.0 line rewrote the agent, so read the migration guide when coming from 0.x.
Where is Browser Use based?
The terms of service say Browser Use is headquartered in the United States and are governed by California law. The privacy policy names Browser Use, Inc. as the responsible entity.
Verdict
Browser Use is worth evaluating if you want to script an AI model driving a real browser from Python and accept that the default setup sends page content to hosted services.
It is open source under MIT, actively released (last push on 19 November 2026, release 1.0.6 on 16 November 2026) and supports a local Chrome with a local Ollama model. The trade-offs: no installer, fast-moving releases that the README says can break things, safeguards that are not a sandbox, paid cloud defaults, and claims about success rates, stealth and security that rest on the maker's own statements. PromptQuorum has not tested it; try it first on a throwaway profile and a low-risk task.
Sources
- Browser Use website — products and pricing, checked 19 November 2026.
- GitHub repository — repository facts and last push, checked 19 November 2026.
- README — install paths, browser modes, models and FAQ, checked 19 November 2026.
- LICENSE — MIT license, checked 19 November 2026.
- Release 1.0.6 — latest release, published 16 November 2026.
- pyproject.toml — version and Python 3.11+ requirement, checked 19 November 2026.
- Privacy policy — responsible entity and data collected, checked 19 November 2026.
- Terms of service — headquartered in the United States, California law, checked 19 November 2026.
- Localhosted listing — directory listing with tags, checked 19 November 2026; the listing is not claimed by its owner.
