Thirdwatch MCP docs

Setup

The MCP server speaks the standard Model Context Protocol over HTTP. Any MCP-compatible client works. Configurations for the three most common ones:

Claude Desktop

Edit your Claude Desktop config file, restart the app, and "thirdwatch" appears in the Tools menu.

// claude_desktop_config.json
// macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
// Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "thirdwatch": {
      "url": "https://mcp.thirdwatch.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_THIRDWATCH_KEY"
      }
    }
  }
}

Cursor

Drop the JSON below in either the project-scoped or global config. Cursor picks it up on next reload.

// .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)
{
  "mcpServers": {
    "thirdwatch": {
      "url": "https://mcp.thirdwatch.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_THIRDWATCH_KEY"
      }
    }
  }
}

Claude Code

One command. Use --scope user to make it available across every project.

# Add to Claude Code (user scope, available everywhere)
claude mcp add --transport http --scope user thirdwatch \
  https://mcp.thirdwatch.dev/mcp \
  --header "Authorization: Bearer YOUR_THIRDWATCH_KEY"

# Verify it loaded
claude mcp list

For other clients, point them at https://mcp.thirdwatch.dev/mcp with an Authorization: Bearer … header. The server speaks JSON-RPC 2.0; all standard MCP methods work (tools/list, tools/call).

Auth & keys

Every request needs your Thirdwatch key in the Authorization header.

• Generate a key by signing up at /pricing. Each account gets one key on creation; you can rotate it from the dashboard.
• Treat keys like passwords. Don't commit them; load from environment variables or a secrets manager.
• Rate limits are per-key. Default is 60 requests per minute per key, plenty for interactive agents. Talk to support for higher limits.
• Compromised key? Rotate from the dashboard or email support. The old key dies immediately on rotation.

Credits & billing

• 1 credit ≈ $0.01 on the Free tier. Larger packs lower the effective rate.
• Each tool call costs 1–26 credits depending on how many sources it queries — see the tool reference below.
• Failed calls aren't billed. If the call errors before returning data, no credits are deducted.
• Credits never expire. Buy them once, spend them when you need them.
• Check balance any time with the free get_account_info tool, or the dashboard.

Pricing detail and credit packs: /pricing.

Calling tools (raw protocol)

If you're building your own MCP client (no Claude/Cursor), the wire format is JSON-RPC 2.0. Discover tools with tools/list then call them with tools/call.

// JSON-RPC over HTTP — what your client sends to https://mcp.thirdwatch.dev/mcp
{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "search_jobs",
    "arguments": {
      "query": "senior product designer",
      "location": "Bangalore",
      "max_results": 25
    }
  },
  "id": 1
}

The response wraps the result in an MCP content block. Each tool's input shape is published in its inputSchema (visible via tools/list); see the reference below for what each one does.

Tool reference

All 14 tools, grouped by cluster. Credit costs are minimum–maximum ranges; use estimate_cost to get the exact price for specific arguments.

Errors

Errors come back as standard JSON-RPC error objects. The data field always carries a type tag you can switch on.

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": {
    "code": -32602,
    "message": "Insufficient credits",
    "data": {
      "type": "insufficient_credits",
      "balance": 12,
      "required": 18
    }
  }
}

Example prompts

Drop these into Claude or Cursor with the MCP connected. The agent picks the right tool; you don't need to name it.