Guide · 7 min read
Anvil ships an MCP server. Claude Code, Cursor, and Windsurf can drive your real devices — list, tap, screenshot, replay — without leaving the editor. Here's the five-minute setup.
The Model Context Protocol server surfaces Anvil's verb catalog as MCP tools — one tool per verb, namespaced anvil__*. Your AI IDE can call anvil__devices_list, anvil__spec_run, anvil__tap, anvil__screenshot, and so on — the same surface the CLI gives you, wrapped as MCP.
It also exposes four resources: anvil://runs/recent, anvil://devices, anvil://specs, and anvil://findings — so the IDE can answer "what broke on the last iPhone 16 Pro run?" without you having to pipe JSON.
The MCP server is part of the CLI. If you have @anvil/cli installed globally, you already have it.
# Install npm i -g @anvil/cli # Login so the MCP server has an API key anvil login --api-key ak_live_... # Verify anvil mcp --print-config
anvil mcp --print-config emits a JSON snippet you can paste into any MCP-aware client.
Add to your Claude Code settings (either the project's .claude/settings.json or the global one):
{
"mcpServers": {
"anvil": {
"command": "anvil",
"args": ["mcp"],
"env": {
"ANVIL_API_KEY": "${ANVIL_API_KEY}"
}
}
}
}Reload the agent. The anvil__* tools appear in the tool list. Ask Claude to "run the smoke suite against my iPhone and tell me what broke" — it dispatches, waits, reads the replay, summarizes.
Cursor and Windsurf use the same MCP manifest format. Drop the snippet above into their settings panel (both have an "MCP servers" section) and save. No restart needed on Cursor; Windsurf requires a reload.
Both IDEs honor the ANVIL_API_KEY environment variable — so your team can share a config file without leaking keys.
The MCP server runs verbs against whichever device matches the call — including BYOD devices on your desk. For safety, three rails are on by default: (1) destructive verbs (uninstall, factoryReset, clearKeychain) require --allow-destructive on the session; (2) calls to devices inproduction class fail without an explicit --prod flag; (3) rate-limited to 120 verbs/minute per client.
All three rails are configurable per-org. The defaults err on the side of "an AI will do something unexpected and we'd rather it be recoverable."
Ready to try it?
Free for 100 runs/month. No credit card.