Skip to main content
OVRSEA runs a Model Context Protocol server. An assistant connected to it can query shipments, quotations, invoices and documents in natural language, and act on them, without building an integration.

​
Endpoint

​
Connect

The API Integration page in Hermès installs Claude and Cursor in one click, and shows the current steps for the others in your own language. It is the canonical source: if an assistant changes its settings layout, that page is updated first. The steps below are the same flows, for reference.
Add the server, then authenticate:
--scope user registers the server for every project; without it, it exists only in the directory you ran the command in.On Claude Desktop, add it via Settings → Connectors → Add custom connector with the endpoint.

​
Authentication

The server uses OAuth, not the API token used by the REST API. The assistant opens an OVRSEA sign-in on first connection, and every later call carries that user’s own identity as a bearer token. Two consequences worth planning around:
  • A connection is personal. The assistant sees exactly what that user sees in Hermès. Two colleagues connecting the same assistant get different views if their Hermès access differs.
  • There is no shared service credential. For unattended, machine-to-machine access, use the REST API with a token instead.

​
What the assistant can do

Once connected, the assistant discovers the tools itself — asking it what it can do is always more current than a list written here. Reading. Shipments and their tracking, cargo, pricing, documents and activity; quotations and their PDFs; invoices; customs clearances; BL, AWB and CMR instructions; saved addresses and contacts; notifications, tasks and pending actions; reporting data; negotiated rates on a lane. Writing. Requesting a quote, accepting or declining one, opening a shipment, sending a booking request, BL instructions, booking details, cargo-ready and target delivery dates, addresses and contacts, renaming, and cancelling a shipment. Agent actions marks which actions are available here.

​
Before you automate

Every tool is published with the standard MCP annotations describing its effect: readOnlyHint for reads, destructiveHint for changes that cannot be undone, idempotentHint for calls that are safe to repeat. Interactive assistants normally use them to decide when to ask the user first. An unattended client has no such prompt, so gate on the action itself rather than on the annotation alone.
The MCP specification states that tool annotations are “not guaranteed to provide a faithful description of tool behavior”. Several tools carry destructiveHint — among them opening a shipment, sending a booking request, accepting a quote, replacing or deleting BL instructions, and cancelling a shipment. declineQuotation does not, yet reopening a declined quotation brings the quotation back without its prices.