For agents, not humans

The MCP server
that sells travel.

AI agents need a transaction layer. We are it. Live since 2024. Twelve MCP tools public. Authenticated reservations across 1.5M+ properties. Primary MCP partner for ChatGPT, Claude, Cursor, Windsurf, and every agent OS shipping in 2026.

mcp.reservations.comJSON RPC 2.06 PUBLIC TOOLS99.98% UPTIME$2B+ ANNUAL GMV

Every agent needs a way to actually reserve.

When a user asks their AI to reserve a hotel, the agent runs into a wall. Property data lives behind OTA paywalls. Rates require credentials. Availability needs live calls. Payment needs tokenization. Cancellation needs policy lookup. None of this is solved at the model layer.

The agent has three options. Scrape, which breaks, gets rate limited, and exposes the user to bad data. Redirect to a human flow, which defeats the point of having an agent. Integrate directly with a reservation layer, which is what reservations.ai provides.

We built the MCP endpoint first, before MCP was a published spec. We have been running production MCP traffic since Anthropic released the protocol. That lead time matters: every other travel API retrofitting MCP today is adapting an existing REST surface, not designing for agents first.

Travel was the first commerce vertical where MCP actually closed loops end to end. Everyone else is still wiring up search.MCP Ecosystem Partner · Q1 2026

From intent to confirmation in under a second.

A typical agent reservation flow hits six tools in sequence, all through the MCP endpoint. End to end latency is sub second for the common path. No redirects, no OAuth dances, no hand off to a human form.

· RESERVATION FLOW · typical path
STEP 01agent calls tools/list discover schema, zero credentials
STEP 02agent calls search_properties destination, dates, guest count
STEP 03agent calls get_rates live rates with taxes and fees
STEP 04agent calls check_availability real time, locks soft hold
STEP 05agent calls create_booking atomic, tokenized payment
STEP 06agent calls confirm_booking two phase commit finalize
DONEconfirmation number issued, PNR returned AEC refunded 100%

The two phase commit pattern (create then confirm) lets the agent present the full reservation details to the user for approval before money moves. For fully autonomous agents operating under user pre authorization, both phases collapse into one call.

No UI. No humans in the loop.

Agents register themselves. POST to /agents/register, receive credentials, load the AEC pool, start transacting. The endpoint has no browser surface, no OAuth consent screen, no onboarding funnel. It is built for machines.

POST · agents/registermachine only
# Agent registers itself, no human in the loop
curl -X POST https://mcp.reservations.com/agents/register \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "claude-booking-1.0",
    "principal": "user_7f3e...",
    "pre_auth": {
      "max_usd": 2500,
      "scope": ["hotel"],
      "expires_at": "2026-06-01T00:00:00Z"
    },
    "aec_pool": 500.00
  }'

# Response — credentials and pool balance
{
  "agent_token": "ra_live_01HZ8K3...",
  "aec_balance": 500.00,
  "rate_limit": { "rpm": 1200 },
  "manifest_url": "/manifest.json",
  "status": "active"
}

Every registered agent is bound to a user principal, a spend cap, a scope (hotels, cars, activities), and an expiration. The agent can transact freely within these bounds. Exceed the cap or the expiration, the token is revoked, no exceptions.

Key rotation is automatic. Agent tokens rotate every 90 days by default, or immediately on user principal logout. The MCP client library handles rotation transparently, no code changes needed.

We are the first reservation layer that an agent can actually self serve. The rest of travel tech still assumes a human is behind the keyboard.Neil Valentine · Chief Strategy Officer

Plug into every surface shipping an agent.

Native MCP clients get drop in support. Non MCP clients get the same tools via OpenAI tool schema or direct REST. Webhooks cover the async surface for every state change.

Claude Desktop
LIVE

Add reservations.ai to your MCP server config. Claude discovers every tool on launch. Works with Claude Code, Claude Desktop, and Claude mobile apps.

mcp_config.json
Setup guide
ChatGPT
LIVE

Registered as an official ChatGPT travel app. Available to all ChatGPT users through the apps surface. Handles voice, text, and image inputs.

apps.openai.com/reservations
Install in ChatGPT
Cursor
LIVE

MCP server config drops in. Use from Cursor Composer or inline chat. Ideal for travel startups building reservation UX, letting Cursor agent generate and test integrations.

.cursorrules
Install in Cursor
Windsurf
LIVE

Codeium Windsurf MCP support ships out of the box. Use for building travel apps or for travel scenarios inside agent workflows.

windsurf_config.json
Setup guide
Custom agents
LIVE

Any agent that speaks JSON RPC 2.0 can connect. Use the Python or Node SDK, or raw HTTP. Full OpenAPI 3.1 spec available for non MCP frameworks.

pip install reservations-ai
Read the SDK docs

Six tools public. Zero credentials.

The public MCP endpoint at mcp.reservations.com/public/mcp exposes 6 tools to any client without authentication. Search, quote, check availability, review policy, and do every read heavy operation a discovery agent needs.

MethodToolPurposeTier
GETtools/listDiscover MCP schema and capabilitiesPublic
POSTsearch_propertiesQuery 1.5M+ properties by destination, dates, filtersPublic
POSTget_ratesLive rates with all taxes, fees, and inclusionsPublic
POSTcheck_availabilityReal time availability with soft hold optionPublic
POSTget_policyCancellation, change, and refund policy fetchPublic
POSTget_amenitiesFull amenity list, images, room typesPublic
POSTcreate_bookingAtomic reservation with payment tokenizationAuth
POSTconfirm_bookingTwo phase commit finalizeAuth
POSTcancel_bookingCancellation respecting property policyAuth
GETget_bookingRetrieve reservation record by confirmation numberAuth
POSTmodify_bookingChange dates, guest count, room typeAuth

Authenticated tools require a Reservation Partner tier token. Enterprise tools (voice routing, custom rate overrides, regional inventory pinning) are exposed only to Enterprise partners with a named integration engineer.

In active development

Hotels and lodging are the live tool surface today. Flights, car rentals, activities and tours, and show and event tickets are being built on the same MCP surface, the same auth model, and the same two phase reservation pattern. They are not callable yet. Tool names are provisional. Anything not returned by tools/list is not yet shipped.

VerticalPlanned toolsStatus
flightssearch_flights · get_fares · create_flight_order · confirm_flight_orderIn dev
car_rentalssearch_cars · get_car_rates · create_car_bookingIn dev
activitiessearch_activities · get_activity_details · create_activity_bookingIn dev
show_ticketssearch_events · get_seat_map · create_ticket_orderIn dev

The direction is one MCP surface for all of travel: lodging, air, ground, activities, and live events, on the same credentials and AEC metering the hotel tools use today.

Already shipping an agent?
Plug in in 30 seconds.

READ AGENTS.MD →

Agents that close pay nothing.

AEC (Autonomous Execution Credits) meter every tool call. A $500 minimum pool opens the door. Credits never expire. When a reservation completes, every AEC consumed in that path refunds to the pool at 100%. Net cost for a successful agent: zero, plus the negotiated revenue share on reservation GMV.

Agents that browse without converting still pay for the compute they used. That keeps the endpoint fast for everyone by pricing the behavior correctly. Full AEC rate card published on the pricing page.

Revenue share is variable, negotiated at eSign onboarding based on integration depth, reservation volume, channel mix, and geography. Standalone agent integrations typically settle inside industry affiliate ranges. Enterprise white label terms go lower in exchange for volume commits and multi year terms.