> ## Documentation Index
> Fetch the complete documentation index at: https://apidoc.ovrsea.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Introduction

> How an AI agent reaches an OVRSEA account

An AI agent can reach an OVRSEA account three ways. They differ in who drives the call, how the
caller signs in, and what the agent is allowed to change.

## MCP endpoint

```
https://mcp.ovrsea.com/mcp
```

Point an MCP-capable assistant at that URL and sign in with an OVRSEA account. See
[MCP server](https://apidoc.ovrsea.com/ai/mcp-server.md) for the per-client configuration.

## The three channels

| | Driven by | Sign-in | Can change |
| - | - | - | - |
| **MCP server** | your AI assistant | OAuth, the user's own account | most of the agent actions |
| **REST API** | your code | API token, from the Hermès API Integration page | purchase orders only |
| **Ask Zach** | a person, in Hermès | already signed in | the agent actions |

> [card] [MCP server](https://apidoc.ovrsea.com/ai/mcp-server.md)
>
> Connect Claude, ChatGPT, Copilot, Cursor or Mistral.

> [card] [REST API](https://apidoc.ovrsea.com/introduction.md)
>
> Read shipments, invoices, documents and tracking from your own code.

> [card] [Ask Zach](https://apidoc.ovrsea.com/ai/ask-zach.md)
>
> The assistant already inside Hermès, for the people using it.

## Permissions

Every channel resolves permissions server-side, against the identity its credential carries, and
an agent cannot reach anything that identity cannot already reach. This is enforced by the
credential rather than by the agent.

Which identity that is differs by channel. The MCP server and Ask Zach both act as the signed-in
person, so they see exactly what that colleague sees in Hermès and no more. A REST token is issued
for the account by an owner or admin, so it is not scoped to one person — treat it as an account
credential and store it accordingly.

## What an agent may change

[Agent actions](https://apidoc.ovrsea.com/ai/agent-actions.md) lists every action an agent can perform, generated from the API
schema, and marks which are exposed as MCP tools.

Reading never requires confirmation. Changing does, and how depends on the channel:

* **REST API** — no approval step. A call with a valid token performs the write immediately, and
  your code is the control.
* **MCP server** — every tool carries the standard MCP annotations (`readOnlyHint`,
  `destructiveHint`, `idempotentHint`). Interactive assistants normally use them to decide when to
  ask the user first; an unattended client has no such prompt. Read
  [Before you automate](https://apidoc.ovrsea.com/ai/mcp-server.md) before relying on them.
* **Ask Zach** — the assistant states what it is about to change and waits for the person to agree,
  on a card for the most common actions and in the conversation for the rest.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.