# Browser-Use

> Vision-first browser automation built for AI agents.

[Canonical HTML page](https://top-mcps.com/mcp/browser-use) · [server.json](https://top-mcps.com/mcp/browser-use.json) · [methodology](https://top-mcps.com/about/methodology)

## Install

### Claude Desktop — `claude_desktop_config.json`

Paste under mcpServers. Fully quit and reopen Claude after editing.

```json
{
  "mcpServers": {
    "browser-use": {
      "command": "uvx",
      "args": [
        "browser-use-mcp"
      ]
    }
  }
}
```

### Claude Code — `CLI or .mcp.json`

Run from your repo. Commit .mcp.json to share with your team.

```shell
claude mcp add browser-use -- uvx browser-use-mcp
```

### Cursor — `.cursor/mcp.json`

Global path: ~/.cursor/mcp.json. Reload window after editing.

```json
{
  "mcpServers": {
    "browser-use": {
      "command": "uvx",
      "args": [
        "browser-use-mcp"
      ]
    }
  }
}
```

### VS Code — `.vscode/mcp.json`

VS Code uses the "servers" key (not "mcpServers").

```jsonc
{
  "servers": {
    "browser-use": {
      "command": "uvx",
      "args": [
        "browser-use-mcp"
      ]
    }
  }
}
```

### Windsurf — `~/.codeium/windsurf/mcp_config.json`

Open via Cascade → hammer icon → Configure.

```json
{
  "mcpServers": {
    "browser-use": {
      "command": "uvx",
      "args": [
        "browser-use-mcp"
      ]
    }
  }
}
```

### Cline — `cline_mcp_settings.json`

Open via the Cline sidebar → MCP Servers → Edit.

```json
{
  "mcpServers": {
    "browser-use": {
      "command": "uvx",
      "args": [
        "browser-use-mcp"
      ]
    }
  }
}
```

### Continue — `~/.continue/config.json`

Continue uses modelContextProtocolServers with a transport block.

```json
{
  "experimental": {
    "modelContextProtocolServers": [
      {
        "transport": {
          "type": "stdio",
          "command": "uvx",
          "args": [
            "browser-use-mcp"
          ]
        }
      }
    ]
  }
}
```

### Codex CLI — `~/.codex/config.toml`

Codex uses TOML. Each server is a [mcp_servers.<name>] subtable.

```shell
# ~/.codex/config.toml
[mcp_servers.browser-use]
command = "uvx"
args = [
  "browser-use-mcp",
]
```

### Zed — `~/.config/zed/settings.json`

Zed calls them "context_servers". Settings live-reload on save.

```jsonc
{
  "context_servers": {
    "browser-use": {
      "command": {
        "path": "uvx",
        "args": [
          "browser-use-mcp"
        ]
      }
    }
  }
}
```

### ChatGPT — `ChatGPT → Apps directory`

Browser-Use doesn't ship a hosted HTTPS endpoint today. ChatGPT supports remote MCP servers only — to use this server in ChatGPT you'll need to deploy it to a public HTTPS URL first (e.g. via Cloudflare Workers or Vercel) or wait for an official remote build.

```none

```

## At a glance

- **Maintainer:** Browser-Use
- **Transport:** stdio
- **Auth model:** None
- **Required secrets:** None
- **Supported clients:** Claude, Cursor, Windsurf, Any MCP-compatible client, Chromium (auto-installed)
- **License:** MIT
- **Language:** Python
- **Latest version:** latest
- **Last verified:** 2026-05-27
- **GitHub stars:** 95,798 (fetched 2026-06-02T11:55:52.423Z)
- **Score:** 88/100 (rubric 2026-04 — see https://top-mcps.com/about/methodology)
- **Source:** https://github.com/browser-use/browser-use

## Security & scope

- **Access scope:** exec
- **Sandbox:** Spawns a Chromium browser. Sandbox by running the MCP inside a container with no host filesystem mounts and a network policy that only allows the domains the workflow intends to visit.
- **Gotchas:**
  - Browser sees whatever the model navigates to — prompt injection from page content is the headline risk.
  - Headed mode shows everything on the screen; do not point it at personal accounts in shared sessions.
  - Cookies and storage persist per profile dir; isolate per workflow.

## Quick answer

**What it does.** Drives a real Chromium browser via a vision-first action API. Exposes navigation, click-by-element-description, fill, screenshot, and structured page-content extraction.

**Best for:**
- Visual web automation
- Sites with shifting selectors
- Click-by-description workflows
- Form fill with screenshot verification
- Multi-step agentic browsing

**Not for:**
- Deterministic test suites
- High-volume scraping
- Headless-mandatory production deploys

## Description

Browser-Use is an open-source browser-automation MCP designed from the ground up for AI agents. Instead of asking the model to write Playwright selectors, the MCP gives the model a rendered snapshot of the page and a high-level action API — click, type, scroll. The model decides what to do based on what it sees, not on brittle CSS paths.

## Why it matters

Selector-based browser automation breaks every time a vendor updates their HTML. Vision-first automation lets the agent adapt to UI changes the way a person does.

## Key features

- Vision-first action API
- Chromium via Playwright under the hood
- Screenshot + DOM dual context
- Element-by-description targeting
- Multi-tab navigation

## FAQ

### How is it different from Playwright?

Playwright wants selectors. Browser-Use wants a screenshot. The model "sees" the page and decides what to click — robust against UI changes that would break selector-based scripts.

### Headless mode?

Supported but discouraged during development — the screenshots are how the model orients. Run with headed mode while iterating, switch to headless for production.

### Does it sandbox?

Not by default. For untrusted page content (which is most agent browsing), run inside a container with no host filesystem mounts and a network egress allowlist.

### Performance on big pages?

Screenshot + DOM extraction is slower than pure DOM ops — expect ~1–3 sec per action vs sub-second for Playwright. Trade-off for resilience.

## Changelog

- **2026-05-27** — Refreshed install snippets and fact sheet; verified for 2026.
- **2025-03-05** — Initial directory listing.