# MCP

> An MCP endpoint that exposes the platform to external AI clients, and an MCP client that lets the platform call external servers - both under the caller's own authority

Source: https://aipril.dev/docs/platform/integrations/mcp

## Inbound: your AI tools reach the platform

An MCP endpoint exposes the platform's operations to an external client - Claude Desktop, Cursor, or
an agent somebody wrote. Over eighty registered tools in thirteen groups: the query engine, datasets,
applications, scanning, memory, knowledge, exploration, skills, pages, processes, ledger,
spreadsheets, dashboards.

Two properties make that safe to expose:

- **Authority is the caller's.** Row policies and column masks are applied by the
  [engine](https://aipril.dev/docs/engine/under-the-hood.md) for the identity behind the call, and process reads run under
  an actor-scoped token and not a service identity. An external client is not a privileged door.
- **Every tool declares a risk classification**, so a write-capable tool called by an unattended agent
  goes through the same [gate](https://aipril.dev/docs/harness/invocation.md) as one called in chat.

## Outbound: the platform reaches your MCP servers

The platform is also an MCP client. An organisation administrator installs a third-party MCP server as
a proxy skill: configuration encrypted at rest, calls forwarded over JSON-RPC, the endpoint hardened
against server-side request forgery, and every call written to the audit trail. See
[Skills](https://aipril.dev/docs/harness/skills.md).

## Rate limits

Per IP, alongside the platform's other limits. See
[Models and providers](https://aipril.dev/docs/harness/models-and-providers.md).

## Detail

The full tool-plane description - groups, the connector bridge, per-caller authority, progressive
disclosure - is on [Tools and MCP](https://aipril.dev/docs/harness/tools-and-mcp.md).