run_browser_agentLABSCO SUMMARY
Two tools, both of which take minutes and cost model calls.
This is not a set of click primitives — you state the outcome and an agent works out the steps, which is why the transport is a long-lived daemon and why the step cap matters. Pointing it at your own Chrome profile is what makes tasks behind a login possible at all.
What it is
Browser-use wrapped as a server. You describe the outcome and a browser agent, driven by an LLM you configure, navigates, fills forms, clicks and extracts what you asked for. It runs as a persistent HTTP daemon rather than over stdio, because a browser task takes minutes and stdio connections drop mid-run.
What you get
- A browser task stated in plain English and carried out end to end — typically a minute or two of navigating, clicking and reading
- A deep research run that fans out across several searches and synthesises the result, rather than returning a list of links
- A live web UI on the daemon's own port showing the task list, progress, logs and server health, plus a dashboard with task history and skills
- Your own browser used when it matters: a persistent Chrome profile so logins and cookies survive, or a connection to a Chrome you already have open over its debugging port
- The model behind the agent chosen by you — Anthropic, OpenAI, Google, Azure, Groq, DeepSeek, Cerebras, Ollama, Bedrock, OpenRouter and others — with vision on by default and a step ceiling per task
- A skills system in beta, off unless you enable it, which records what worked and replays it
Requirements
Python with uv, and Chromium installed through Playwright. An API key for whichever LLM drives the agent — this is the real cost, since every step is a model call; the key goes in the environment rather than the config file. The daemon binds to 127.0.0.1 on port 8383 and clients connect to /mcp over streamable HTTP, with mcp-remote bridging clients that only speak stdio. Settings live in a config file under your home directory, with environment variables taking precedence over it. Defaults worth knowing: the browser runs headless, a task is capped at 20 steps, and a research run at 5 searches. Exposing it beyond localhost means setting an auth token.
Setup effort
Build from source — git clone https://github.com/Saik0s/mcp-browser-use.git
USE CASES
01
Getting something off a site with no API
Describe what you want in a sentence and let the agent navigate and extract it, instead of scripting selectors that break next week.
02
Doing a task behind a login
Run it against a persistent Chrome profile, or attach to the Chrome you are already signed into, so the session is already there.
03
Researching a question across many sources
Let it run several searches and synthesise them into an answer, rather than handing back a page of links to read yourself.
04
Filling in a form you would otherwise do by hand
State the values and the destination and let the agent work through the fields, watching progress in the web UI while it runs.
2 tools
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
run_deep_researchBEFORE YOU INSTALL
What to know first
Tools you get
2 tools when we connected. run_browser_agent takes a plain-language task and does the clicking, filling and extracting itself, up to agent.max_steps (default 20). mcp_server_browser_use is the Python package name, not a tool.
Requires
- uv
- An API key for whichever LLM provider you configure, in your environment — GEMINI_API_KEY for the default Google provider, ANTHROPIC_API_KEY for Anthropic, and so on
CONTEXT COST
Light Standing cost
Rated by the size of the 2 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
Connect your client
This one is built from source rather than installed with a single command.
Server setup
TERMINAL
git clone https://github.com/Saik0s/mcp-browser-use.git
cd mcp-browser-use
uv sync
uv run playwright install chromiumIn the same terminal, set the API key for the LLM that drives the browser agent (GEMINI_API_KEY for the default Google provider, ANTHROPIC_API_KEY for Anthropic): export GEMINI_API_KEY=<YOUR_GEMINI_API_KEY>
TERMINAL
uv run mcp-server-browser-use serverMCP Clients
Claude Code — plugin
CLAUDE CODE — PLUGIN · IN CLIENT
/plugin marketplace add Saik0s/mcp-browser-use
/plugin install mcp-browser-use@browser-useAny MCP client — HTTP
ANY MCP CLIENT — HTTP · JSON
{
"mcpServers": {
"browser-use": {
"type": "streamable-http",
"url": "http://localhost:8383/mcp"
}
}
}Clients spell the "type" value differently: Claude Code and VS Code use "http", Cline uses "streamableHttp", and Roo Code uses "streamable-http". Claude Desktop does not accept a URL entry in its config file.
Any MCP client — mcp-remote
ANY MCP CLIENT — MCP-REMOTE · JSON
{
"mcpServers": {
"browser-use": {
"command": "npx",
"args": [
"mcp-remote",
"http://localhost:8383/mcp"
]
}
}
}
