Skip to main content
Create project

Connect Claude or Cursor

Sheepit ships an MCP (Model Context Protocol) server that exposes ~37 tools — campaigns, flags, experiments, dashboards, insights queries — directly to Claude Desktop, Cursor, Codex, and any other MCP-capable client. Once connected, the LLM can read your data and ship changes for you without leaving the IDE.

Three commands. After this you can ask Claude to “create a dashboard for last week's signups” or “turn on the new_pricing_page flag for users in the dogfooders group” and it'll just do it.

1 Sign in from the terminal
A browser tab opens, you confirm the project, and a long-lived secret key lands at ~/.sheepit/credentials.json. The MCP server picks it up automatically.
bash
npx @sheepit-ai/cli login
2 Install the MCP server in Claude Desktop / Cursor
Auto-detects which MCP-capable clients you have installed and writes thesheepit entry into each one's config file. Restart the client after to see the tools.
bash
npx @sheepit-ai/mcp install --yes
Manual config (if your client isn't auto-detected)
claude_desktop_config.json
{
  "mcpServers": {
    "sheepit": {
      "command": "npx",
      "args": ["-y", "@sheepit-ai/mcp", "serve"]
    }
  }
}
3 Ask Claude to do something
The MCP server exposes ~37 tools — campaigns, dashboards, flags, experiments, insights queries. Try one:
text
Ask Claude:
  "Create a dashboard that tracks signups and errors
   over the last 30 days for my Sheepit project."

Other prompts that work well: “What are my top traffic sources this week?”, “Create a feature flag called new_pricing_page off by default, then turn it on for users in the dogfooders group.”

What is MCP?

Model Context Protocol is the open standard for letting an LLM call tools on a remote server. Instead of pasting API keys into a system prompt, the IDE spawns a local process (the MCP server) that holds your credentials and exposes a structured tool catalog. Claude / Cursor / Codex all speak the same protocol, so the same @sheepit-ai/mcp install works across every client.

Where are my credentials stored?

npx @sheepit-ai/cli login runs a PKCE OAuth round-trip in your browser, mints a lp_sec_* secret key scoped to the project you select, and writes it to ~/.sheepit/credentials.json (mode 0600). The MCP server reads that file on startup. To switch projects, run sheepit login again. To rotate the key, revoke it from Settings → API Keys.

Troubleshooting

  • Claude doesn't see the tools after install. Quit Claude Desktop completely (⌘Q on macOS — closing the window isn't enough) and re-open it. MCP servers are spawned at app launch.
  • @sheepit-ai/cli login hangs. Check that nothing else is bound to port 8765 (the loopback callback port). The CLI retries up to 5 ports automatically.
  • @sheepit-ai/mcp install says “no clients detected”. Either Claude Desktop / Cursor / Codex aren't installed, or their config files are in non-default locations. Use the manual JSON snippet (under step 2 above) and paste it into your client's MCP config yourself.
  • Tools fail with “No project context”. Your secret key is tied to one project. To work against a different project, run sheepit login again and pick the new one.

Related

  • Quick Start — install @sheepit-ai/react / @sheepit-ai/server in your app and start emitting events.
  • API Keys — view, rotate, or revoke the key MCP is using.