# Scout [Health: Active]

**Category:** 📂 Browser Automation  
**Repository:** https://github.com/felixgeelhaar/scout  
**GitHub Stars:** 14  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/scout-2

## Description
Browser automation MCP server with 84 tools. Pure CDP, DOM diffing, vision hybrid.

## Claude Desktop Quick Installation
No generated config for this server: it installs through its own setup steps.

The project's own setup steps (from https://github.com/felixgeelhaar/scout), copied verbatim. Run them in order; they are not an mcpServers config.

1. Install:

```sh
brew install --cask klarlabs-studio/tap/scout-mcp
```

2. Add to Claude Code:

```sh
claude mcp add scout -- scout mcp serve
```

3. Install:

```sh
curl -fsSL https://raw.githubusercontent.com/klarlabs-studio/scout/main/install.sh | bash
```

4. Install:

```sh
go install go.klarlabs.de/scout/cmd/scout@latest
```

5. Add to Claude Code:

```sh
claude mcp add scout -- scout mcp serve           # Claude Code
```

## Documentation

## What the Scout MCP server does

The Scout MCP server exposes browser automation to MCP clients through one statically linked executable. It can open pages, inspect structured browser state, interact with controls, extract content, capture screenshots or PDFs, and manage multiple tabs. The same project also provides a CLI, a conversational browser UI, and a Go library, but the MCP command is the agent-facing entry point.

Its browser engine communicates directly with Chrome through the DevTools Protocol. This makes it suitable for tasks such as navigating websites, filling forms, collecting table data, checking application state, and diagnosing failed requests. The tool set also covers iframes, cookies, dialogs, authentication walls, console errors, web vitals, and file uploads.

## How it works

Start the server with `scout mcp serve` and configure an MCP client to launch that command over standard input and output. The agent then calls browser tools within a session. Read-only operations such as observation, extraction, and screenshots carry read-only MCP annotations, while action tools are marked with additional metadata for approval handling.

Page observation can use regular DOM data, DOM differences, token budgets, readable text, Markdown, or accessibility information. `hybrid_observe` and coordinate-based lookup add a vision-oriented path when DOM selectors are insufficient. Browser state can be retained while navigating, opening tabs, switching tabs, or recording a session.

## Setup and configuration

Install the binary with Homebrew, the project’s install script, or `go install`. A basic MCP configuration uses the `scout` executable with `mcp` and `serve` as its arguments. No Node.js or Python runtime is needed for the MCP server.

The browser can run headless or visibly. Use the `configure` tool to change that mode during a session. Private and loopback addresses are blocked unless explicitly enabled with `configure(allow_private_ips: true)` or the `SCOUT_ALLOW_PRIVATE_IPS=1` startup environment variable. This option is intended for trusted local-development workflows.

## Tools and capabilities

The Scout MCP server includes tools for:

- Navigation, clicking, typing, hovering, dragging, scrolling, and event dispatch.
- Form discovery and semantic form filling, including checkbox and radio state.
- Element, table, page-text, Markdown, accessibility, and automatic extraction.
- Screenshots, annotated screenshots, PDFs, screen recordings, and traces.
- Tab and iframe management, cookies, network capture, failed-request inspection, and console diagnostics.
- Framework detection, SPA waiting, component or application state inspection, and web vitals.
- Recording and replaying JSON playbooks, batch execution, selector suggestions, readiness checks, and session history.

Screen recordings are returned as file paths. With `ffmpeg` available, Scout can encode WebM or MP4; otherwise it returns raw frames and an encoding list for offline processing.

## Limitations and notes

Screen capture polling typically produces around 10–15 frames per second and is capped at 30. Localhost and private-network access requires an explicit opt-in.

_Full upstream README: https://allmcps.com/mcp/scout-2/readme_