An MCP browser server gives your AI agent a browser. The agent calls tools like navigate, click, type and screenshot through the Model Context Protocol, and the server translates those calls into real browser actions. No Playwright code, no CSS selectors, no custom function definitions. The LLM discovers available tools, decides which ones to call and reads the results.
TL;DR: An MCP browser server exposes browser actions as tools an LLM can call. Add one config block to Claude Code, Cursor or any MCP client, and your agent can browse the web. Clawbrowser's built-in MCP server adds what most servers skip: managed fingerprints, proxy routing and anti-detection so the agent works on real websites, not just demo pages.
What an MCP Browser Server Does
MCP (Model Context Protocol) standardizes how LLMs connect to external tools. An MCP browser server is an MCP server that exposes browser actions as callable tools over JSON-RPC 2.0.
Here is what happens when your agent navigates to a page:
- The LLM decides it needs to visit a URL
- It sends an MCP tool call:
navigate(url="https://example.com") - The MCP server receives the call and translates it into a CDP command to the browser
- The browser loads the page
- The server returns the result (page title, content snapshot, status) to the LLM
- The LLM reads the result and decides what to do next
LLM ← MCP (JSON-RPC) → MCP Server ← CDP (WebSocket) → Browser
tool calls clawctl browser commands ClawbrowserThe LLM never touches the browser directly. It works entirely through the MCP tool interface: describe what you want, call the right tool, read the result.
For a deeper comparison of the two protocols in that stack, see CDP vs MCP: When to Use Which for AI Browser Control.
The Problem Most MCP Browser Servers Ignore
Every MCP browser server on the market connects to some version of Chrome or Chromium. The protocol part works fine. The browser part is where agents break.
A stock Chromium instance connected through an MCP server leaks the same automation signals that get detected through any other method:
navigator.webdriveris set totrue- The User-Agent string says "HeadlessChrome"
- Canvas and WebGL fingerprints match known automation hashes
- CDP protocol artifacts are visible to page-side JavaScript
- No proxy routing means the IP mismatches the browser locale
The LLM makes the right decisions. It navigates correctly, clicks the right elements, fills forms accurately. Then the site serves a CAPTCHA wall or a blank page because the browser underneath failed the fingerprint check.
This is not an MCP problem. It is a browser problem. The MCP server faithfully executed the tool calls. The browser was the one that got flagged.
For the full detection model (20+ signals anti-bot systems check), see Browser Fingerprinting Explained.
Setting Up Clawbrowser's MCP Server
Clawbrowser is a local, free, Chromium-based anti-detect browser with managed fingerprints, proxy routing, native CDP and a built-in MCP server. The MCP server connects to Clawbrowser's own patched Chromium engine where fingerprints are managed at the source level across 20+ surfaces and CDP signal suppression happens before page-side JavaScript can detect it. The MCP server exposes all of this through standard MCP tools.
Install Clawbrowser
Paste the install prompt from clawbrowser.ai into your AI agent (Claude Code, Cursor, Codex). The agent runs the installer automatically.
Or install manually with the CLI:
clawctl install --json
clawctl config set api-key <your-key>Connect to Your Agent
One command for all supported agents:
clawctl install --agent all --jsonThis configures the MCP server for Claude Code, Cursor, Codex, Gemini CLI and other supported agents automatically.
Manual configuration (add to your agent's MCP config):
{
"mcpServers": {
"clawbrowser": {
"command": "clawctl",
"args": ["mcp", "serve"]That is the entire setup. No API keys for the MCP server itself, no cloud accounts, no Docker containers. The server runs locally and the agent discovers available browser tools on startup.
Verify the Connection
Ask your agent to test the browser:
Open https://pixelscan.net/ip and tell me the page title.The agent should call the navigate tool, load the page and return the title. Pixelscan checks your browser fingerprint and IP — if it shows a clean result, your stealth setup is working. The fingerprint and proxy management are already active.
Tools the MCP Server Exposes
When an LLM connects to Clawbrowser's MCP server, it discovers these browser tools:
| Tool | What it does | Example |
|---|---|---|
| navigate | Load a URL in the browser | navigate(url="https://news.ycombinator.com") |
| click | Click an element on the page | click(selector="#login-button") |
| type | Type text into an input field | type(selector="#search", text="MCP browser") |
| screenshot | Capture the current page as an image | screenshot() |
| extract | Pull structured data from the page | extract(selector=".post-title") |
| scroll | Scroll the page or a specific element | scroll(direction="down") |
| tab_open | Open a new browser tab | tab_open(url="https://example.com") |
| tab_switch | Switch to a different tab | tab_switch(tab_id=2) |
| tab_close | Close a tab | tab_close(tab_id=2) |
The LLM chains these tools together based on the task. "Go to Hacker News and find posts about MCP" becomes a sequence of navigate, scroll, extract and possibly screenshot calls, all decided by the LLM in real time.
Each tool call runs inside a Clawbrowser profile. The profile holds a consistent fingerprint, proxy assignment and session state (cookies, localStorage). If the agent logs into a site in one session, the login persists in the next session on the same profile.
Real Workflow: Agent Extracts Pricing Data
Here is what a real MCP workflow looks like. The user asks the agent to compare pricing across three SaaS tools.
User prompt:
Go to linear.app/pricing, notion.so/pricing and height.app/pricing.
Extract the plan names and prices from each. Summarize in a table.What the agent does (MCP tool calls):
1. navigate(url="https://linear.app/pricing")
2. screenshot() # Read the page visually
3. extract(selector=".pricing-tier") # Pull plan data
4. tab_open(url="https://notion.so/pricing")
5. screenshot()What the agent returns:
| Tool | Free | Pro | Business |
|---|---|---|---|
| Linear | $0 | $8/user/mo | $14/user/mo |
| Notion | $0 | $10/user/mo | $15/user/mo |
| Height | $0 | $8.50/user/mo | Custom |
The agent decided the navigation order, which selectors to try, how to interpret the page content and how to format the output. The MCP server handled the browser. Clawbrowser handled the fingerprints and proxy routing. Each pricing page saw a browser that looked like a real user, not an automation script.
MCP Browser Server Comparison
Not all MCP browser servers are the same. Here is how the main options compare:
| Feature | Clawbrowser | Playwright MCP | Browserbase | Puppeteer MCP |
|---|---|---|---|---|
| Runs locally | Yes | Yes | No (cloud) | Yes |
| Free | Yes | Yes | Paid | Yes |
| Anti-detection | Engine-level (20+ surfaces) | None | Cloud-managed | None |
| Proxy routing | Built-in | None | Included | None |
| Session persistence | Profile-bound | None | API-managed | None |
| CDP fallback | Same browser instance | Separate setup | Separate endpoint | Separate setup |
| Fingerprint management | Automatic | None | Cloud-managed | None |
| Agent integrations | Claude Code, Cursor, Codex, Gemini CLI | Any MCP client | Any MCP client | Any MCP client |
| Best for | Real-world automation on protected sites | Dev/testing on unprotected pages | Parallel cloud execution | Lightweight scripting |
Playwright MCP and Puppeteer MCP are good starting points for development. They work on sites without anti-bot protection. For production agents that hit real websites behind Cloudflare, DataDome or Akamai, you need a browser that handles fingerprints and proxy routing at the engine level.
For detailed comparisons with specific competitors, see Clawbrowser vs Browserbase and Clawbrowser vs Browser Use.
When MCP Alone Is Not Enough
MCP gives the LLM high-level browser tools. That covers most agent workflows. But some tasks need lower-level control that MCP tools do not expose:
- Network interception: blocking tracking scripts, capturing API responses, modifying request headers
- JavaScript injection: running custom scripts before page load
- Performance-critical loops: scraping thousands of pages where LLM inference latency per action is too slow
- Custom wait conditions: waiting for specific DOM mutations or network idle states
For these, drop to CDP. Clawbrowser exposes a CDP endpoint on every profile alongside the MCP server. Connect Playwright or Puppeteer to the same browser instance the MCP server uses:
from playwright.async_api import async_playwright
async with async_playwright() as p:
browser = await p.chromium.connect_over_cdp("http://127.0.0.1:9222")
page = browser.contexts[0].pages[0]Same fingerprints, same proxy routing, same session state. Different interface. Use MCP for the parts where the LLM reasons about what to do. Use CDP for the parts where you need deterministic, fast control.
For a full tutorial building a CDP-based agent, see How to Build an AI Agent That Browses the Web.
FAQ
What is an MCP browser server?
An MCP browser server is a program that exposes browser actions (navigate, click, type, screenshot) as tools an LLM can call through the Model Context Protocol. The LLM sends tool calls over JSON-RPC 2.0, the server translates them into browser commands and returns the results. It is the bridge between an AI agent and a web browser.
Do I need an MCP browser server or can I just use Playwright?
If your agent is an LLM that decides actions based on page content, use an MCP browser server. The LLM discovers tools automatically and calls them through a standard interface. If your automation follows fixed steps with no LLM reasoning, use Playwright directly over CDP. If you need both patterns in the same project, Clawbrowser supports both protocols on the same browser instance.
Can MCP browser servers handle anti-bot detection?
The MCP protocol itself does nothing for anti-detection. What matters is the browser the server controls. Playwright MCP and Puppeteer MCP connect to stock Chromium, which leaks automation signals. Clawbrowser's MCP server connects to a patched Chromium with managed fingerprints across 20+ surfaces, built-in proxy routing and CDP signal suppression at the source level.
How is Clawbrowser's MCP server different from Browserbase?
Clawbrowser runs locally on your machine for free. Browserbase runs in the cloud with usage-based pricing. Clawbrowser manages fingerprints at the Chromium engine level. Browserbase handles anti-detection through cloud infrastructure. Both expose MCP tools, but Clawbrowser also gives you a CDP endpoint on the same browser instance for hybrid workflows.
Which AI agents work with MCP browser servers?
Any agent that supports MCP as a client. Clawbrowser has native one-command integration with Claude Code, Cursor, Codex, Gemini CLI and Kilo Code. Other MCP-compatible agents work with manual configuration.
Give Your Agent a Browser
Install Clawbrowser from clawbrowser.ai by pasting the install prompt into Claude Code, Cursor or your preferred AI agent.
Run clawctl install --agent all --json to configure the MCP server for all supported agents. Your LLM gets browser tools immediately: navigate, click, type, screenshot, extract. Every tool call runs through a browser with managed fingerprints, proxy routing and anti-detection built in.
No cloud accounts. No usage meters. No stealth plugins. One config block and your agent browses the real web.
Continue exploring
Ask AI how Clawbrowser helps
Keep reading
Related articles
Session Persistence for AI Browser Agents: Why Your Agent Keeps Losing State
AI browser agents lose session state between runs because standard tools don't persist cookies and fingerprints together. Here's how to fix it with named browser profiles.
Read article →
CDP vs MCP: When to Use Which for AI Browser Control
CDP controls the browser. MCP connects the AI. Learn when to use Chrome DevTools Protocol vs Model Context Protocol for AI agent browser automation.
Read article →