@onegent/mcp-server

Travel Booking Agent — Book restaurants, hotels, flights, and activities through an AI agent that navigates real booking sites.

Onegent — AI books your trip end-to-end.

This is a Model Context Protocol server that exposes Onegent's trip-booking execution engine to AI clients — Claude Desktop, Claude.ai, ChatGPT Apps, or any other MCP-compatible LLM host.

Instead of giving the LLM generic web-browsing tools and hoping it figures out a booking flow, this server hands it six purpose-built tools that map directly to Onegent's production REST API. The agent navigates OpenTable / Resy / Booking.com / Expedia / Viator on your behalf and stops before charging your card — you confirm the final submit.

What it can book

| Tool | What it does | Typical completion | |---|---|---| | book_restaurant | Dining reservations on OpenTable, Resy, direct sites | 30–120 s | | book_hotel | Rooms on Booking.com, Expedia, Hotels.com | 60–180 s | | book_flight (preview) | Airfare via Expedia / airline sites | 90–240 s | | book_activity (preview) | Tours, tickets, attractions on Viator / GetYourGuide | 60–180 s | | get_job_status | Poll a booking's progress | — | | get_job_audit | See what the agent observed (for diagnosing failures) | — |

All bookings run asynchronously. book_* tools return a jobId; the LLM should poll get_job_status until the status is terminal.

Install

Claude Desktop

Add this block to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "onegent": {
      "command": "npx",
      "args": ["-y", "@onegent/mcp-server"],
      "env": {
        "ONEGENT_API_KEY": "ogk_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Restart Claude Desktop. A hammer icon should appear in the composer; clicking it shows "Travel Booking Agent" with 6 tools.

Claude.ai (remote MCP — coming in 0.2.x)

OAuth-backed remote MCP support is on the roadmap. Until then use the Desktop path above.

ChatGPT Apps

See docs/integrations/chatgpt-apps.md in the main repo.

Other MCP clients

Any host that launches stdio MCP servers via command/args works. Set ONEGENT_API_KEY in the server's env.

Get an API key

1. Visit https://onegent.one/developers (launching with v0.2.36+ — email [email protected] for early access)
2. Sign in and create a key scoped to the scenarios you need
3. Keys are prefixed ogk_live_ for production or ogk_test_ for sandbox

Environment variables

| Variable | Required | Default | Description | |---|---|---|---| | ONEGENT_API_KEY | yes | — | Format: ogk_(live\|test)_xxxxxxxx…. Treat like a secret. | | ONEGENT_API_BASE_URL | no | https://onegent.one/api/v1 | Override for self-hosted / staging |

Usage (sample LLM turn)

User: Book Carbone in NYC for 2 people on April 28 at 7pm. My email is [email protected], phone +1-415-555-0123.

LLM: (calls book_restaurant with restaurant_name="Carbone", city="New York", date="2026-04-28", time="19:00", covers=2, profile={...})

→ Returns jobId: abc-123, status queued.

LLM: (after 20 seconds, calls get_job_status with jobId="abc-123")

→ status running, provider opentable.

LLM: (after 30 more seconds, polls again)

→ status done, confirmationCode XYZ-789.

LLM: "Your table is booked. Confirmation: XYZ-789."

Troubleshooting

ONEGENT_API_KEY env var is required The key isn't reaching the server. Check the env block in your MCP client config is a proper JSON object, not a string.

401 invalid_api_key The key format is right (ogk_live_…) but it doesn't match any active key. Regenerate at /developers.

400 zod validation error The LLM called a tool with bad args. The error message lists which fields are wrong — usually date format (must be YYYY-MM-DD) or missing profile/profileId.

Job stuck in queued for minutes Onegent's executor queue may be backed up. Call get_job_audit for details, or email [email protected] with the jobId.

paused_payment is expected Onegent never submits CVV on its own — by design. The user must confirm the final charge in the Onegent web app. This is a safety feature.

Source access

The package source lives in a private repo. For source access (eval / contribution / security audit), email [email protected].

Report bugs

Email [email protected] with:

• The jobId from the failing run (if applicable)
• Your MCP client (Claude Desktop / ChatGPT Apps / other) + version
• A short description of what you expected vs. what happened

License

MIT