
Agent Infra Browser MCP Server: Setup for Claude Code, Cursor and Other Apps
@agent-infra/mcp-server-browserbytedance/ui-tars-desktop/tree/HEAD/packages/agent-infra/mcp-servers/browserLets Claude, Cursor and other assistants drive a real browser: open pages, click, type and read content.
This server gives Claude, Cursor and other assistants a real browser they can drive: open pages, read them as text or Markdown, click, type, switch tabs and take screenshots. It runs on your computer with npx and needs no account. A vision option adds screenshot-based clicking. It can submit forms and run scripts, so use it on sites you trust.
Good for
- Developers who want the assistant to test a web page or fill a form
- Researchers who need pages that only load in a real browser
- Anyone automating a repetitive web task with an assistant
Works in
12 should work
Tap an app to see how to set it up and any known problems.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Client does not support this transport.
Client does not support this transport.
Install Agent Infra Browser MCP
Claude Code
- 1
Copy this prompt.
Add the Agent Infra Browser MCP server to Claude Code. Run this command: claude mcp add agent-infra-browser -- npx -y @agent-infra/mcp-server-browser Then check that it loaded and tell me if I need to sign in or restart.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js.
Built from the package data we hold. Not yet tested by us.
Tools
Opens a web address
Loads a URL in the browser.
Try: “Open https://news.ycombinator.com.”
Network ·
browser_navigateReads the page as Markdown
Returns the content of the current page as Markdown.
Try: “Give me the current page as Markdown.”
Network ·
browser_get_markdownReads the page as plain text
Returns the text content of the current page.
Try: “What text is on this page?”
Network ·
browser_get_textLists every link on the page
Returns all links on the current page.
Try: “List all links on this page.”
Network ·
browser_read_linksLists the things you can click
Returns the clickable, hoverable and selectable elements with numbers the other tools use. The tool says to call it once per page.
Try: “What can I click on this page?”
Network ·
browser_get_clickable_elementsTakes a screenshot
Captures the page or one element, with options for size, full page and highlighting the element.
Try: “Take a full-page screenshot.”
Network ·
browser_screenshotClicks an element
Clicks the element with the number from the clickable list.
Try: “Click the "Sign in" button.”
Write ·
browser_clickTypes into a form field
Fills an input by number or CSS selector, and can clear it first.
Try: “Type "wireless mouse" into the search box.”
Write ·
browser_form_input_fillPicks an option in a list
Chooses a value in a dropdown by number or selector.
Try: “Select "Large" in the size dropdown.”
Write ·
browser_selectHovers over an element
Moves the pointer over an element, for menus that open on hover.
Try: “Hover over the Products menu.”
Write ·
browser_hoverPresses a keyboard key
Presses a key such as Enter, Tab or an arrow key.
Try: “Press Enter.”
Write ·
browser_press_keyScrolls the page
Scrolls by a number of pixels, or to the bottom if you give none.
Try: “Scroll to the bottom.”
Write ·
browser_scrollGoes back one page
Goes to the previous page in history.
Try: “Go back.”
Write ·
browser_go_backGoes forward one page
Goes to the next page in history.
Try: “Go forward.”
Write ·
browser_go_forwardOpens a new tab
Opens a URL in a new tab.
Try: “Open https://example.com in a new tab.”
Write ·
browser_new_tabLists open tabs
Returns the open tabs.
Try: “Which tabs are open?”
Network ·
browser_tab_listSwitches to another tab
Switches to the tab with the number you give.
Try: “Switch to tab 2.”
Write ·
browser_switch_tabCloses the current tab
Closes the tab you are on.
Try: “Close this tab.”
Write ·
browser_close_tabCloses the browser
Closes the whole browser when the task is done.
Try: “Close the browser.”
Write ·
browser_closeLists downloaded files
Returns the files the browser has downloaded.
Try: “What files did the browser download?”
Network ·
browser_get_download_listRuns JavaScript on the page
Executes JavaScript you or the assistant write in the browser console of the current page.
Try: “Run a script that returns the page title.”
Runs code ·
browser_evaluateTakes a screenshot for vision mode
Captures the screen so a model that reads images can pick where to click. Only available when you start the server with --vision.
Try: “Capture the screen so you can see the page.”
Network ·
browser_vision_screen_captureClicks at screen coordinates
Clicks at an x and y position from a screenshot. Only available when you start the server with --vision.
Try: “Click at the position of the green button in the screenshot.”
Write ·
browser_vision_screen_click
Read from the source code
Permissions
high risk · 3 things to watch
High risk
The assistant drives a real browser that can open any address, fill in and submit forms, click buttons and run JavaScript on the page, so it can act on sites where you are signed in. Use it on sites you trust, and watch the browser window while it works.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It can run JavaScript on any page, with no read-only mode
browser_evaluate executes script in the open page, and click, type and form tools act on live sites. On a site you are signed in to, that can change or delete data. Keep approval prompts on and use a browser profile without your logins. github.com
Pages can try to steer the assistant
Text on a web page is returned to the assistant. Treat pages from unknown sites as untrusted. github.com
Vision tools are off by default
The two screen-click tools only exist when you start the server with --vision. github.com
Security advisories
No published security advisories as of Nov 1, 2026.
Technical details
| Turning on read-only | No read-only mode. Click, type and script tools are always available. |
|---|---|
| Talks to | Any website the assistant opens |
| Sign-in and keys | No account or key. Any logins live in the browser it opens; the README does not say how the browser profile is handled. |
Questions
Does it open a visible browser?
Yes by default. The --headless option hides the window. Other options pick Chrome, Edge or Firefox, a browser path, or attach to a browser you already run.
Do I need an API key?
No. It runs on your computer with no account.
What does vision mode do?
The --vision option adds two tools that work from screenshots and screen positions, for pages where the normal element list is not enough. Without it the assistant uses the page structure.
Can it run on a server with no screen?
The README shows running it on a virtual display by setting the DISPLAY variable, and it can run as a web server with the --port option.
Does it fill in forms and click buttons for me?
Yes, those are core tools. It can also run JavaScript on the page, so check what it plans on sites you are signed in to.
Information
updated Nov 2, 2026
- Publisher
- Community
- Package
@agent-infra/mcp-server-browser- Registry
- npm
- Transport
- stdio, SSE, Streamable HTTP
- Auth
- None
- Version
- 1.0.0
- License
- MIT
- Last commit
- Oct 9, 2026
- Open issues
- 358
- Listed in
- Official MCP registry
Sources
Docs, issue trackers and discussions this page is built from. Issues, quotes and findings above link to their own source.
Show all 4Hide
- Browser server README · github.com
- Tool registrations · github.com
- Command-line options · github.com
- Server manifest · github.com