Claude Desktop Integration Guide

The status

Claude Desktop connects to Orbyt Intelligence today. The config below is the working form. Earlier on June 9, 2026 this page said the opposite, because our MCP server answered the handshake with a protocol version no official client accepts. We fixed that the same day. The history is below, because we promised to keep this page honest.

One requirement up front: the hosted MCP server needs an Intelligence key on the Pro plan or higher. Free-tier keys get a 403 with an upgrade pointer. There is no trial, so a Pro subscription bills from the day you start it.

Two Orbyt surfaces, two different protocols

Orbyt Labs has two products with two different machine interfaces. They are not interchangeable.

The /api/mcp/ prefix on the Jobs API is naming, not protocol. Those 23 endpoints speak plain REST. There is no MCP server for your Jobs pipeline today, so Claude Desktop cannot reach your pipeline. If you want an AI assistant on your pipeline right now, use the ChatGPT GPT Actions guide, Apple Shortcuts, or Zapier. Those work.

The Intelligence MCP endpoint at https://www.orbytlabs.ai/api/v1/intelligence/mcp is a real MCP server, and Claude Desktop can now finish the handshake with it.

What was broken, and what we changed

Claude Desktop, and the mcp-remote bridge it uses for a remote server, are built on the official MCP SDK. During initialize, the SDK checks the protocol version the server returns. Our server used to return 2026-05-10, an internal lock date that is not a real MCP spec revision. The SDK accepts only official revisions, so the connection died with:

Server's protocol version is not supported: 2026-05-10

We reproduced that on June 9, 2026 by running the latest mcp-remote against our own server, then shipped the fix the same day. The server now negotiates protocol versions the way the MCP spec describes: if your client requests 2025-11-25, 2025-06-18, 2025-03-26, or 2024-11-05, the server echoes it back. Anything else gets our newest supported version, 2025-11-25. Official clients complete the handshake.

The config

This is the working claude_desktop_config.json shape for connecting Claude Desktop to the Intelligence MCP endpoint through the mcp-remote bridge.

The file lives at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "orbyt-intelligence": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://www.orbytlabs.ai/api/v1/intelligence/mcp",
        "--transport",
        "http-only",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": {
        "AUTH_HEADER": "Bearer intelligence_your_key_here"
      }
    }
  }
}

Two details worth knowing:

• The header value goes through the env block because some clients split command arguments on spaces. Authorization:${AUTH_HEADER} has no space, so it survives intact.
• mcp-remote needs Node 18 or newer on your machine.

Intelligence API keys are generated from the Intelligence dashboard. The MCP server requires Pro or higher; see pricing for plans. There is no trial, and the free tier does not include MCP.

Restart Claude Desktop after saving the config. You are connected when you ask "What does a Senior AI Engineer in San Francisco make?" and Claude answers with a number and a citation.

You can also talk to the endpoint directly

The Intelligence MCP server answers raw JSON-RPC too. If you are building your own agent, or you just want to see real data without a client in the middle, this works:

curl -X POST https://www.orbytlabs.ai/api/v1/intelligence/mcp \
  -H "Authorization: Bearer intelligence_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "analyze_compensation",
      "arguments": { "role": "ai-engineer", "city": "san-francisco" }
    }
  }'

You get back a decision-ready answer with a citation and a request id. Six tools are available: list_capabilities, analyze_compensation, analyze_skills, analyze_market, discover_roles_and_cities, and find_adjacent_opportunities. Full details on the MCP page.

Your pipeline, other assistants

The Intelligence MCP connection covers salary and labor market data. Your Jobs pipeline is a separate surface, and it is reachable by machine through these routes:

• ChatGPT GPT Actions: imports our trimmed OpenAPI spec and gets all 23 pipeline operations. Verified working.
• Apple Shortcuts: Siri briefings and quick-add via plain REST calls.
• Zapier / Make / n8n: real-time webhooks on pipeline events.
• Anything that speaks HTTP: the full REST surface is documented at /openapi.yaml with Bearer ext_ token auth.

A native MCP server for the Jobs pipeline is on the roadmap. When it ships, this page gets the config the same day.

{
  "mcpServers": {
    "orbyt-intelligence": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://www.orbytlabs.ai/api/v1/intelligence/mcp",
        "--transport",
        "http-only",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": {
        "AUTH_HEADER": "Bearer intelligence_your_key_here"
      }
    }
  }
}