MCP server

The Ultra MCP server turns every /api/v1/* operation into a tool any LLM can call. A travel-agent AI inside Claude Code can read an itinerary, check supplier availability, append a flight, and record a confirmed booking — all without writing a line of HTTP code. That's the point: the operator describes the trip in natural language, the agent does the work against Ultra, the human approves.

Drop the server into any MCP-compatible client — Claude Code, Cursor, Windsurf, Zed, custom clients — and every endpoint becomes a tool the model can reach for.

Why MCP

The MCP server is spec-driven — it loads the live OpenAPI document at boot and registers one tool per operation. When we ship a new endpoint, the next time you restart the MCP server it appears as a new tool. No MCP code changes, no client updates. Your agent grows as the platform grows.

This is the structural bet: agents and humans need to share the same operations on the same data, and the spec is the contract that keeps both honest. The CLI ships from the same spec for the same reason — humans at terminals and agents in chat windows hit the same surface, with the same auth and the same guarantees.

Install

npm install -g @ultra-network/mcp

Or invoke without installing via the config below.

Configure Claude Code

Pick one of three config locations, depending on scope:

• User-global: ~/.claude.json — server available in every project
• Project-local: .mcp.json in the project root — server available only when Claude Code runs from that project (good for keeping a project-scoped ULTRA_API_KEY)
• Claude Desktop: claude_desktop_config.json (location varies by OS — see Anthropic's docs)

The shape is the same everywhere:

{
  "mcpServers": {
    "ultra": {
      "command": "npx",
      "args": ["-y", "@ultra-network/mcp"],
      "env": {
        "ULTRA_API_KEY": "ulk_yourkey"
      }
    }
  }
}

Restart Claude Code. Type /mcp to confirm the server connected. Once it has, the LLM can call tools like list_trips, create_booking, append_trip_item — every v1 operation, exposed as a tool with a name and JSON-schema input.

Configure Cursor / Windsurf / Zed

Each editor has its own MCP config UI. The shape of the entry is the same — a command + args + env block. Point command at your local tsx and args at the MCP server entry file.

Environment

What the LLM sees

Tool names are derived from operationId, snake-cased:

Each tool's inputSchema is derived from the OpenAPI schema — path params, query params, and request body all surface at the top level so the LLM knows what to send.

Restricting tool surface

To expose only a subset of operations (e.g. read-only):

{
  "mcpServers": {
    "ultra-readonly": {
      "command": "npx",
      "args": ["-y", "@ultra-network/mcp"],
      "env": {
        "ULTRA_API_KEY": "ulk_yourkey",
        "ULTRA_API_TAGS": "Trips,Suppliers"
      }
    }
  }
}

For finer-grained control (no createBooking, no updateTripItem), use a key with restricted scopes — see authentication § scopes.

Troubleshooting

tools/list returns 0 tools — ULTRA_API_SPEC may be wrong. Check the URL returns a valid OpenAPI document with non-empty paths.

Every tool call returns 401 — ULTRA_API_KEY is missing or invalid. Boot the server with the var set; the server warns to stderr if it's missing.

Tool calls return 400 "Write operations require an org-bound API key" — you're using a cross-org master key. Use an org-bound key for write operations.

MCP server can't be found by Claude Code — /mcp shows the server status. If it's stuck on "connecting", check ~/Library/Logs/Claude/mcp.log (macOS) for the actual error.

Package

@ultra-network/mcp on npm. MIT-licensed.

Next

• API reference — what each tool actually does on the server
• CLI — same operations, shell-friendly