Claude Desktop / Claude Code / Cursor (MCP)

Zero-install: point your MCP client at the hosted Floe endpoint.

Configure

One line, any client:

npx -y add-mcp https://mcp.floelabs.xyz/mcp

Claude Code:

claude mcp add --transport http floe https://mcp.floelabs.xyz/mcp \
  --header "Authorization: Bearer YOUR_FLOE_KEY"

Claude Desktop / Cursor: add to claude_desktop_config.json or .cursor/mcp.json:

{
  "mcpServers": {
    "floe": {
      "type": "http",
      "url": "https://mcp.floelabs.xyz/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_FLOE_KEY"
      }
    }
  }
}

Which key? Both formats work, and they unlock different tools. A developer key (floe_live_…) drives the lifecycle tools — create agents, mint and rotate keys, set budgets, read balances and usage, register webhooks. An agent key (floe_…) drives the runtime tools — x402_pay, cost estimates against real credit, spend limits, allowlist. Run one server entry per key if you want both. With no key at all, get_markets, check_x402_url, and search_floe_docs still work. See API Keys.

What you get

All 73 Floe MCP tools — agent lifecycle, key minting, spend controls, cost preflight, x402 payment execution, funding instructions, usage analytics, webhooks, and docs search. See the MCP Server reference for the full tool catalog.

Narrow the set with scope params — ?read_only=true or ?features=spend,pricing,payments on the endpoint URL.

Local install (optional)

FLOE_API_KEY=floe_YOUR_AGENT_KEY npx -y @floelabs/mcp-server --stdio

Use this if you want the client to spawn the server locally (handy for development / debugging). --stdio matters: without it the server starts an HTTP listener instead of speaking MCP on stdout.

Example

See floe-cookbook/mcp-demo for a zero-code Claude Desktop config.