Browser Use MCP: give your AI agent a real browser

What is a browser agent?

A browser agent is an AI model that operates a web browser to finish a task. It opens pages, reads them, clicks, types, scrolls and fills in forms until the job is done, then reports what it found. You describe the outcome in plain words - "list the plans on these three pricing pages", "fill in this form with test data" - instead of writing selectors and scripts that break when a page changes.

Browser Use is an open-source agent built for this: it is MIT-licensed, runs on any LLM, and works locally or on Browser Use Cloud, the company's hosted service. Browser Use MCP is how an AI assistant reaches it. The Model Context Protocol (MCP) lets Claude, Cursor, Codex and other clients call tools on a server, and the FlyMy.AI MCP includes Browser Use among its 1,500+ tools. The assistant you already use can hand a task to a cloud browser, follow it, and use the result in the rest of its work.

Why give your AI agent a real browser

Search and fetch tools cover a lot, but much of the web is built for people rather than for APIs: pricing pages that render with JavaScript, search filters, multi-step forms and dashboards behind a login. A browser reaches whatever a person can reach.

• It works on pages, not endpoints. The agent uses a site the way a visitor does, so it needs no API, no scraping rules and no selectors.
• It runs in the cloud. Browser Use Cloud hosts the browser, so nothing runs on your laptop and the task keeps going while your assistant does other work.
• You can watch and step in. The person who started a task gets a private live view and can take over for a login, a 2FA code or a CAPTCHA.
• Logins carry over. A saved browser profile keeps a login for the next task.
• It leaves evidence. Each task returns its output, status and cost. Recordings, screenshots and downloads can be saved as files, and now a read-only replay can be shared.

What's new in Browser Use MCP

This release moves FlyMy.AI's Browser Use tools to Browser Use Cloud v4, with gpt-6-luna as the default model. Each change is a parameter or an action your assistant can call, and the platform log lists every one of them.

Our own test sessions on 11 November 2026; times and costs as Browser Use Cloud reported them.

How to set up Browser Use MCP in a minute

You add one MCP server to your assistant, sign in and ask. There is no Browser Use account to create and no key to copy.

Open example.com in a browser on bu-ultrafast and tell me the page title.

Prefer to let the assistant do the setup? Paste this into any assistant:

Read https://flymy.ai/skill.md and https://flymy.ai/mcp/browser-use.md. Help me connect Browser Use MCP to this assistant. Then help me run this task: Open news.ycombinator.com in a real browser, collect the top 10 story titles and links, and show me the recording.

Pick the right model: bu-ultrafast, bu-fast or gpt-6-luna

The model decides how fast the agent moves and how carefully it reads each page. New sessions use gpt-6-luna; the two Ultrafast presets are built for speed.

• bu-ultrafastfastest
  Short, interactive flows. A form, a few clicks, a quick read. Our tests: a page title in 4 s, a seven-field form in 11 s.
• bu-fastpreset
  Quick, with more reasoning. Thinks more per step than bu-ultrafast, for flows with a few more decisions in them. No thinking_level, like bu-ultrafast.
• gpt-6-lunadefault
  Economical and capable. Takes every thinking level from disabled to max, for research, comparisons and harder pages. Our test: a judged two-step read in 16 s.

The Ultrafast presets take no thinking_level and stay fixed for the session, so choose them when you start the task; they run without the managed proxy unless you set proxy_country_code. On gpt-6-luna, thinking_level goes from disabled to max; use xhigh or max for pages that need careful reading, such as a long comparison.

Compare three project-management tools in a browser with thinking_level max.

Steer a browser task while it runs

Long tasks rarely go exactly to plan: halfway through, you want one more field, a different filter or a second page. Until this release you waited for the task to end and started again. Now send_task reaches a task while it is still running.

• Queued (default). The instruction goes into Browser Use's session queue and the task takes it up next. The response carries a queued_message_id, and wait_for_browser_task waits until the queued instruction has run.
• Interrupt. With interrupt: true the task cancels its current work and runs your instruction now.
• After it finishes. A follow-up continues the same session with its conversation, files and saved profile: in the same browser while it is still open, or in a fresh browser once it has closed. keep_alive: true keeps the browser open up to 30 minutes after the task.

A running task takes only the instruction text. To change model, thinking_level or judge, send the follow-up after the task finishes; a session on an Ultrafast preset keeps its preset.

Start a browser task that compares the pricing pages of Asana, Trello and ClickUp and lists each plan with its monthly price. While it is still running, tell it to also note which plans include a free trial.

Share a read-only replay of a browser session

Sometimes you want to show how the agent got there: to a client, a teammate or a support thread. set_browser_share_link turns on a public, read-only link to the session's replay on Browser Use's page, with the task, the agent's steps and its screenshots, and no control of the browser. Turn it off when you are done; turning it on again restores the same link.

Make a public read-only replay link of my last browser task.

When the review is over, ask: "Turn off the replay link of my last browser task."

Check results with the judge

A browser agent can report success and still miss part of the task. judge: true asks Browser Use's LLM judge to grade the finished task against its instruction. The verdict comes back as judgement, with verdict, reasoning and failure_reason, and is_success follows it: the task counts as a success only when the verdict passes. The judge adds a small model charge to the task.

In the test session shown above, the judge passed the task with this reasoning:

"The agent successfully navigated to the specified URL, identified the page heading, scrolled to the actions table, and extracted the first five actions as requested."

Turn the judge on when nobody watches the run: scheduled checks, unattended jobs and results that feed the next step of a workflow.

Find the opening hours of the Louvre in a browser and have the judge check the answer.

Use cases with ready prompts

Paste any of these into an assistant connected to FlyMy.AI and replace the example site with yours.

1. QA a website on mobile and desktop

Open your pages at phone and desktop sizes, save screenshots and get a list of what looks broken; browser screens can be 320 to 6144 pixels wide. For a whole site, with every link checked and a published report page, use the Website QA agent template.

Open https://example.com in a browser at 390x844 and again at 1440x900. On each size, check the header, the main call to action and the footer, save a PNG screenshot of each view as a file, and give me a short report with the screenshots and what to fix first.

2. Fill in forms in seconds with bu-ultrafast

In our test, bu-ultrafast filled in the seven fields of httpbin.org's sample order form, submitted it and returned the server's echo in 11 seconds, for $0.0024. The agent is told to ask before purchases, sends or anything it cannot undo, so name the final step when you want it done.

On bu-ultrafast, open https://httpbin.org/forms/post, fill in the pizza order form with test data (name Ada Test, size Medium, toppings Bacon and Extra Cheese, delivery 19:30) and submit it. Show me what the server echoed back.

3. Research you can steer

Comparisons take longer, and you often think of one more question halfway through. Start the task, add to it while it runs, and let the default model think harder on dense pages.

In a browser with thinking_level high, find the monthly prices of the paid plans of 1Password, Bitwarden and Dashlane on their own pricing pages. While it runs, ask it to also check whether each one offers a family plan.

4. Show a client what the agent did

Send a read-only replay instead of screenshots pasted into a document: the client sees the task, every step and the answer, and you turn the link off after the review.

Run a browser task that goes through the checkout of https://example.com up to the payment page without paying, then make a public read-only replay link I can send to my client.

5. A weekly check with a verified result

A scheduled agent can run the same browser task every week. With the judge on, a run passes only when the judge agrees, so a missing price or a broken button shows up as a failure instead of a cheerful summary.

Create an agent that every Monday at 9:00 opens https://example.com/pricing in a browser, checks that every plan shows a price and a Sign up button, and runs the browser task with judge: true. Schedule it weekly and run it once now so I can see the result.

Browser Use MCP vs Playwright MCP vs computer use

There are several ways to give an AI agent a browser. They differ in where the browser runs and in who plans each step. The table sticks to what each project documents; the sources are linked below it and were checked on 11 November 2026.

Swipe the table sideways to compare all four.

Sources: Playwright MCP README, Claude computer use tool docs, Browser Use README, FlyMy.AI Browser Use docs.

How to choose. Pick Browser Use MCP when you want a task done in the cloud from one instruction, with a live view, follow-ups and a shareable replay, and without running a browser yourself. Pick Playwright MCP when you want your assistant to drive a local browser step by step, for example while testing your own app. Pick computer use when the task goes beyond the browser into desktop applications.

What it costs

On FlyMy.AI, Browser Use usage on the shared account is billed to your FlyMy.AI balance at Browser Use's own reported price, with 0% markup: the browser time, the agent's model and any proxy traffic. With your own Browser Use key on the connection, the usage is paid on your Browser Use account instead.

Times and costs are Browser Use Cloud's own figures for these sessions. Through an assistant, add a few seconds for the MCP round trip. Prices follow Browser Use's current rates.

• max_cost_usd caps each task's model and search spend, $1 by default; browser time and proxy traffic come on top, and one session never goes past $10.
• A browser kept open with keep_alive keeps costing until it closes, so stop the sessions you no longer need.

Keep reading

• Browser Use MCPEvery action, setup commands and more requests.
• Website QA agent templateA whole site on mobile and desktop, with a report page.
• Browser agents that stay logged inSaved browser profiles for tasks behind a login.
• One MCP, 1,500+ toolsBrowser Use next to search, scraping, Google and Slack.
• Platform logWhat is new on FlyMy.AI, release by release.
• The agent loop is an interpreter. We compiled it.Our study of compiled agent programs.