Sly Browser Agent

Sly is an open Firefox extension and native host for LLM-driven browser automation. It exposes browser control as an MCP server, so it works with any LLM client that supports MCP over stdio. It has no domain blocklist.

Architecture

LLM client <--stdio MCP--> host/mcp-server.js <--TCP--> native host
                                                        <--Firefox extension

The extension provides navigation, screenshots, mouse and keyboard input, accessibility-tree inspection, page text extraction, element search, forms, uploads, console/network diagnostics, window control, and an optional humanized input mode. The execute_code variants are available under host/codemode/ when a sandboxed batch API is desired.

Install

Requirements: Firefox 121+, Node.js 18+, and an MCP-compatible LLM client.

1. Install host dependencies: npm install --prefix host.
2. Open about:debugging#/runtime/this-firefox in Firefox.
3. Choose Load Temporary Add-on and select extension/manifest.json.
4. Copy the generated extension ID and run ./install.sh <extension-id>.
5. Fully restart Firefox so it reloads the native messaging manifest.
6. Configure the client to launch node /absolute/path/to/host/mcp-server.js as an stdio MCP server.

The extension is loaded temporarily by Firefox, so its ID changes if it is loaded again. Re-run install.sh whenever that happens. The Firefox background script is persistent and owns recorder state directly; no hidden recorder tab or offscreen document is needed.

Transcription

The recorder is optional. Its options page accepts an API key, endpoint, and model for any OpenAI-compatible audio transcription API. Defaults target OpenAI's /v1/audio/transcriptions endpoint with whisper-1.

Development

There is no extension build step. Reload the temporary add-on after changing extension/. Install host dependencies with npm install --prefix host, then run npm test from host or node --check against changed JavaScript files.

MIT licensed.