# @roveapi/mcp MCP server for [Rove](https://roveapi.com) — a hosted Playwright API for AI agents. Returns **accessibility trees** instead of screenshots, reducing LLM token consumption by ~77% (26K tokens vs 114K for a typical page). ## Quick Start ```bash npx -y @roveapi/mcp ``` Set your API key via environment variable: ```bash ROVE_API_KEY=rvp_live_... npx -y @roveapi/mcp ``` ## Install in Claude Code ```bash claude mcp add --scope user rove -e ROVE_API_KEY=rvp_live_YOUR_KEY -- npx -y @roveapi/mcp ``` ## Install in Claude Desktop Add to `claude_desktop_config.json`: ```json { "mcpServers": { "rove": { "command": "npx", "args": ["-y", "@roveapi/mcp"], "env": { "ROVE_API_KEY": "rvp_live_YOUR_KEY" } } } } ``` ## Install in Cursor Add to `.cursor/mcp.json`: ```json { "mcpServers": { "rove": { "command": "npx", "args": ["-y", "@roveapi/mcp"], "env": { "ROVE_API_KEY": "rvp_live_YOUR_KEY" } } } } ``` ## Tools | Tool | Description | |------|-------------| | `navigate` | Navigate to a URL. Auto-creates a session. Supports stealth mode and action jitter. | | `interact` | Click or fill actions in the active session. | | `extract_schema` | Extract structured data from a URL using a JSON schema. | | `screenshot` | Take a screenshot (in-session or standalone). | | `get_a11y_tree` | Get the accessibility tree snapshot — the core differentiator. | | `close_session` | Close a browser session and release resources. | ## Why Accessibility Trees? Traditional browser automation returns screenshots that cost ~114K tokens per page. Rove returns structured accessibility trees at ~26K tokens — **77% fewer tokens**, faster responses, and better structured data for LLMs to reason over. ## Configuration | Environment Variable | Default | Description | |---------------------|---------|-------------| | `ROVE_API_KEY` | `rvp_live_demo` | Your Rove API key | | `ROVE_API_BASE_URL` | `https://api.roveapi.com` | API endpoint | ## Get an API Key Sign up at [roveapi.com](https://roveapi.com) — 100 free credits on signup, no card required. ## License MIT