Connect AI agents over MCP
The Model Context Protocol (MCP) is an open standard that lets an AI agent use outside tools. Connect Claude, Cursor or any other MCP client to the platform and your agent can search and price travel, hold offers, read orders and trip data, and prepare bookings that a person on your team approves.
The tools are generated from the same API contract as the REST API, so they behave the same way: the same products, the same rules, the same errors. An agent gets exactly the permissions of the credential it connects with, and never more.
Preview. The hosted MCP endpoint is not open to customers yet. These pages describe the tools and how agents connect, so you can plan your integration now. The endpoint address will be published in the quick-start when it opens. The travel tools will take an API key for agents from your console; signing in to them with OAuth is planned but not available yet.
Toolsets
The tools are split into toolsets by who uses them. Each toolset has its own endpoint, and an agent connected to one cannot see or call the tools of another. The onboarding tools are listed on their own page, first, because that is where an account starts, but they are part of the travel tools: same endpoint, same key.
| Toolset | For | How the agent signs in | Availability |
|---|---|---|---|
| Onboarding tools | Taking your account live: who you are signed in as, the go-live checklist, integration checks and API keys | An API key for agents, on the travel tools' endpoint | Preview |
| Travel tools | Agents that shop, hold and prepare bookings for your customers | An API key for agents | Preview |
| Console tools | Your team's back office: queues, reports, team, webhooks, pricing rules | As you, with OAuth | Sandbox only for now |
| Supplier tools | Suppliers selling through the platform | As you, with OAuth | Sandbox only for now |
| Dining partner tools | Restaurant partners and their linked restaurants | A dining partner API key | Sandbox only for now |
Each reference page lists every tool and resource with what it does, its inputs and an example call. Tools marked Coming soon are documented so you can plan for them; agents cannot list or call them yet.
How agents stay safe
Every tool falls into one of three kinds, and each reference entry says which.
| Kind | Examples | What it takes |
|---|---|---|
| Read | search, price, rules, order and trip reads, travel data | Nothing beyond the credential's own permissions |
| Change that moves no money | a free hold, an enquiry, an upload, a queue claim | Your account must allow agents to make it; otherwise the agent is told it needs approval |
| Money | booking, cancelling | Never done by the agent. It files a request that a person approves |
Money needs a human. Two tools touch money, book_offer and cancel_with_quote, and neither one books or cancels anything. Each gets a live price and then files a request. A person on your team reviews it in the console, sees the exact amount, and approves it in their own signed-in session. The platform then carries it out once, at that amount. The agent cannot approve its own request, and typing "yes" in a chat is not an approval. Payments, refunds and ticket issue are not available to agents at all.
An agent acts with its credential's permissions. No tool takes an account id: the account comes from the credential. An agent connected with an API key has that key's scopes. An agent you connect to the console or supplier tools acts as you, with your current permissions, and loses access when you do.
Supplier content is data. Free text that comes back from suppliers, such as hotel descriptions or fare notes, is passed to the agent as data and never treated as instructions.
Connecting as yourself
The console and supplier tools serve an agent that a signed-in person connects to their own account, using the MCP authorization standard (OAuth 2.1 with PKCE). Your MCP client does the protocol work; you see three things:
Sign-in. The client opens your browser at the console (or the supplier console). Sign in if you are not signed in already.
Consent. The consent screen shows the client's name, whether the platform has verified it, where it will send you back, and every read it asks for. All reads start ticked; untick any you do not want. On the console tools, changes the agent may make are listed separately, every box starts unticked, and ticking one asks you to confirm with your second factor.
Done. Your client receives a short-lived token for that one toolset.
A client without a browser, such as a terminal tool, can show you a code instead. Type it yourself on the console's connect page (/console/connect, or /supplier/connect in the supplier console); no link carries the code.
What a connection is:
One toolset, one account. The token works only on the endpoint it was issued for.
Short and bounded. A connection lasts at most 8 hours and never longer than your own session. Signing out, signing out everywhere, a password change or a disabled user ends it.
Only what you ticked. A read you left unticked is refused. A change runs only if an Owner of your account turned it on for agents and you ticked it at consent.
Never money, never approvals. A connected agent cannot approve a request, approve another agent, change these settings, create a key or move money.
Yours to revoke. See and revoke your connections in Connected agents: Administration → Connected agents in the console, Account → Connected agents in the supplier console. People who manage users see every connection on the account. An administrator can also switch the console tools off for the whole account there.
Letting an agent make changes
On the console tools, a few changes that move no money can be handed to an agent: claiming and releasing a queue item, sending a test webhook, replaying a webhook delivery and escalating a case. Each one runs only when all of these hold, checked again on every call:
An Owner of your account turned that change on in Agent writes (Administration → Connected agents), confirming with a second factor. Every change starts off. Turning one off needs no second factor, so stopping is always quick.
The person who connected the agent ticked that change on the consent screen.
The connection is still live.
Every change an agent makes is recorded with both the agent and the person it acts for.
Limits
Agent calls count against your account's normal API rate limits. On top of them, MCP has a per-minute rate limit and a call budget per account. When the rate limit is reached the answer is RATE_LIMITED, which is retryable: wait and try again. When the budget is spent the answer is AI_BUDGET_EXHAUSTED until the budget window resets.
Errors
Errors come back in the same shape as the REST API, {code, message, traceId, retryable, details}, with nextAction where the API suggests one, so an agent branches on code and retryable the way any client does. API errors pass through unchanged. MCP adds a few codes of its own:
| Code | What it means | What to do |
|---|---|---|
APPROVAL_REQUIRED | The tool makes a change your account has not allowed agents to make | Ask a person to do it, or have an Owner allow it |
PERMISSION_DENIED | The credential, or the reads you ticked, do not cover this call | Reconnect and tick it, or use a key with the scope |
CREDENTIAL_KIND_NOT_SUPPORTED | This toolset needs a different kind of sign-in | Check the toolset's sign-in on its reference page |
AI_FEATURE_DISABLED | Agents, or this toolset, are switched off for your account | Ask an administrator |
RATE_LIMITED | Too many calls in a short time | Retry after a short wait |
AI_BUDGET_EXHAUSTED | The account's agent call budget is spent | Wait for the budget window to reset |
HOLD_EXPIRING | The hold expires too soon for a person to approve a booking on it | Extend the hold with extend_hold, or book the live offer |
PENDING_RECONCILIATION | An earlier attempt reached the supplier and is being checked | Retry later; never book again |
PROPOSAL_NOT_PERSISTED | A booking or cancellation request could not be filed | Retry; nothing was booked or cancelled |
Next steps
Quick-start: connect a client and walk through search, price, hold and a booking request a person approves.
The agent workflow: how an agent takes your company from sign-up to a certified sandbox integration, and the two steps the Owner takes to go live.
Reference: travel · console · supplier · dining partner.
What's new for agents: the tools added or withdrawn in each API version.