CloudBrowser MCP server
Our hosted MCP server adds 20 browser tools to any MCP client. Paste one URL, sign in, and your assistant can browse, click, type and extract data from any website.
https://mcp.cloudbrowser.aiScreen recording of a real Chrome session: an AI agent is asked to “Add a sweater that fits my 30 lb dog to my cart”. It opens demo.vercel.store, closes a popup, searches for “dog sweater”, opens the product, reads the size options, picks 20 – 50 lbs and adds it to the cart.
Works with
Model Context Protocol
Add CloudBrowser once and your assistant gets 20 browser tools — in chat, in your editor or inside your own agents.
https://mcp.cloudbrowser.aiSign in with your CloudBrowser account (OAuth) or send your API token as a Bearer header. Use the URL exactly as shown.
CloudBrowser with the MCP server URL.Developer mode is a ChatGPT beta for eligible paid plans on the web; in Business workspaces an admin sets it up. Menu names vary between ChatGPT releases.
The open-source MCP connector is free. Browser usage requires a CloudBrowser account and a paid plan after the 7-day Basic trial. Get your API token in the dashboard.
How it works
The Model Context Protocol lets an AI client discover tools and call them. CloudBrowser turns each call into real actions in an isolated Chrome — and hands back exactly what the model needs.
ChatGPT, Claude, Cursor, VS Code, Codex or your own agent decides which tool to call.
Checks your account, runs the tool and returns page text, elements or a screenshot.
A real browser with stealth fingerprints, a premium proxy, saved sessions and a live remote desktop.
20 MCP tools
Your assistant decides what to call. CloudBrowser runs each step in a real browser and returns what the agent needs — page content, elements or a screenshot.
Spin up isolated Chrome instances on demand and keep track of every session.
open_browserOpen a stealth Chrome browser in the cloudconnect_to_browserTake direct control of a browserget_browsersList active browsersget_connected_browsersList browsers under controldisconnect_browserRelease control of a browserclose_browserClose a browser when the job is doneLoad any page, including JavaScript-heavy apps, and read what a person would see.
navigate_to_urlGo to any URL and wait for it to loadget_page_contentRead the title, text, headings, links and formsget_elementsExtract elements that match a CSS selectortake_screenshotCapture the viewport or the full pageClick, type and run scripts to search, filter, fill forms and move through flows.
click_elementClick buttons, links and menustype_textType into inputs and formsevaluate_scriptRun JavaScript on the pageWatch live, step in when a human is needed and reuse signed-in sessions.
start_remote_desktopOpen a live view of the browserstop_remote_desktopClose the live viewdetect_captchaSpot reCAPTCHA, hCaptcha and Turnstilebrowser_workspaceShow browsers and sessions in a workspaceget_saved_sessionsList saved sessionsresume_saved_sessionReopen a saved session, still signed inremove_saved_sessionDelete a saved sessionTool reference
Each tool also tells the client whether it only reads or can change something, so assistants can ask you before they act.
| Tool | What it does | Parameters | Type |
|---|---|---|---|
open_browser | Open a stealth Chrome browser in the cloud. | label, keepOpen (300 s), saveSession, recoverSession, browser (Chrome · Chromium · Headless Shell), headless, stealth, args | Manages browsers |
connect_to_browser | Take control of a running browser for page actions. | browserAddress, sessionId | Manages browsers |
get_browsers | List the browsers running in your account. | — | Read-only |
get_connected_browsers | List the browsers this conversation controls. | — | Read-only |
disconnect_browser | Release control of a browser without closing it. | sessionId | Manages browsers |
close_browser | Close a browser so its usage stops. | address | Closes or deletes |
navigate_to_url | Go to a URL and wait until it has loaded. | sessionId, url, waitUntil (load · domcontentloaded · networkidle0 · networkidle2), detectCaptcha | Acts on pages |
get_page_content | Read the page: title, URL, text, headings, links, forms, buttons and images. | sessionId | Read-only |
get_elements | Return the elements that match a CSS selector. | sessionId, selector | Read-only |
take_screenshot | Capture the 1920 × 1080 viewport or the full page. | sessionId, fullPage, type (png · jpeg) | Read-only |
click_element | Click a button, link or menu item. | sessionId, selector | Acts on pages |
type_text | Type into an input or text area. | sessionId, selector, text, clear (true) | Acts on pages |
evaluate_script | Run JavaScript on the page and return the result. | sessionId, script | Acts on pages |
detect_captcha | Spot reCAPTCHA, hCaptcha and Turnstile widgets on the page. | sessionId | Read-only |
start_remote_desktop | Start a live view of a browser so a person can watch or take over. | address | Manages browsers |
stop_remote_desktop | Stop the live view. | address | Manages browsers |
browser_workspace | Show running browsers and saved sessions in a visual workspace, in clients that render MCP Apps. | browserAddress (optional) | Read-only |
get_saved_sessions | List saved sessions by label. | — | Read-only |
resume_saved_session | Reopen a saved session with its cookies and storage — still signed in. | label, requestId | Manages browsers |
remove_saved_session | Delete a saved session. | label | Closes or deletes |
Human in the loop
Every browser can stream to a live remote desktop. When an agent meets a login, a one-time code or a CAPTCHA, you step in — and it carries on from there.
detect_captcha tells the agent when to ask for helpSecurity & auth
Connections are tied to your CloudBrowser account, scoped to browser control and easy to revoke.
Clients that support OAuth register themselves and send you to CloudBrowser to sign in and approve browser control. No token to copy.
Developer tools and agent frameworks can send your API token instead. Keep it in an environment variable on the server, never in client code.
OAuth access tokens expire after 10 minutes and refresh tokens rotate on every use. A grant lasts up to 30 days.
See and revoke connected assistants in your dashboard. Revoking stops further tool calls.
Each browser runs in its own container, and callers only ever see the browsers and sessions of their own account.
Every tool declares whether it is read-only or can change things — clients use these hints to ask before acting.
Hosted or local
The connector is open source. Most people use the hosted server; the npm package is there when a client only speaks stdio.
| Topic | npm package | Hosted (recommended) |
|---|---|---|
| Setup | npx @cloudbrowser/mcp-server | Paste https://mcp.cloudbrowser.ai into your client |
| Transport | stdio, or HTTP with --transport http | Streamable HTTP |
| Sign-in | CLOUDBROWSER_API_TOKEN environment variable | OAuth or Bearer API token |
| Works in web apps | Local desktop and CLI clients only | ChatGPT, Claude and other web clients |
| Updates | Can lag behind the hosted server | Always the newest tools |
| Browsers | The same cloud browsers | Cloud browsers in your CloudBrowser account |
Use it from the OpenAI Agents SDK or Claude Agent SDK The CloudBrowser plugin for ChatGPT
Try asking
Once CloudBrowser is connected, these work as plain prompts in ChatGPT, Claude, Cursor or any MCP client.
“Open the pricing pages of our five main competitors and summarize every plan change in a table.”
navigate_to_urlget_page_content“Find dental clinics in Austin that offer online booking and collect name, website and phone number.”
navigate_to_urlget_elements“Check these three product pages and tell me which one is cheapest right now, with screenshots.”
navigate_to_urltake_screenshot“Walk through the signup flow on our staging site and screenshot each step. Flag anything that breaks.”
click_elementtype_text“Sign in to the supplier portal with my saved session and list last month’s invoices.”
get_saved_sessionsget_elements“Collect the ten latest posts from this blog and turn them into a Markdown digest.”
get_page_contentevaluate_scriptA hosted Model Context Protocol server at https://mcp.cloudbrowser.ai. It gives AI assistants 20 tools to open, drive and read real Chrome browsers running in the cloud.
ChatGPT (developer mode), Claude (custom connectors), Claude Code, Cursor, VS Code, Codex, Gemini CLI, n8n and any other client that supports Streamable HTTP MCP servers. A dedicated CloudBrowser plugin for ChatGPT is coming soon.
Clients that support OAuth let you sign in with your CloudBrowser account and approve browser control. Developer tools can send your API token as an Authorization: Bearer header instead — find your token in the CloudBrowser dashboard.
The open-source MCP connector is free, and it adds no AI inference charge. Browser usage requires a CloudBrowser account and a paid plan after the 7-day Basic trial.
Yes. The @cloudbrowser/mcp-server npm package runs over stdio with your CLOUDBROWSER_API_TOKEN. The npm release can lag behind the hosted server, so we recommend the hosted URL.
Local browser MCP servers drive a browser on your own machine. CloudBrowser runs each browser in an isolated cloud container with stealth fingerprints, a premium proxy, saved sessions and a live remote desktop — and it works in web apps such as ChatGPT and Claude, where a local browser is not available.
detect_captcha flags CAPTCHA widgets so the agent can pause, and start_remote_desktop opens a live view where you can step in. Saved sessions let agents return to sites you are already signed in to.
As many as your plan allows at once: 10 concurrent browsers on Basic and 25 on Premium, each with up to 3 tabs. Custom plans go higher.
Yes. The OpenAI Agents SDK, the Claude Agent SDK, the OpenAI Responses API and other MCP-capable frameworks can call the hosted server with a Bearer token. See the SDK page for ready-to-run examples.
No. Sign in yourself on the live remote desktop and save the session; the agent resumes it later with resume_saved_session. Cookies and storage stay in your CloudBrowser account.
Start with a 7-day free trial. Connect ChatGPT, Claude or your code in minutes.
