# Model Context Protocol (MCP)

MCP is an open standard for connecting AI assistants to external tools and
data sources. Poly speaks it: platform features — the secrets vault today,
more over time — can be exposed to any MCP-compatible client.

## How MCP works

MCP uses JSON-RPC 2.0. A server exposes three primitives:

| Primitive | Description | Example |
| --- | --- | --- |
| Tools | Actions the AI can perform | list vaults, fetch a secret |
| Resources | Read-only data sources | a vault listing, an env export |
| Prompts | Pre-built prompt templates | a setup walkthrough |

The client discovers what a server offers, then calls it on the model's
behalf — the model sees tool descriptions and results, never raw
credentials.

## Where Poly uses MCP

- **Secrets** — the vault is MCP-accessible inside Poly, governed by the
  activity log and leases. See [Secrets & MCP](/docs/secrets-mcp-integration).
- **Assistants** — direct connections from Claude Code and Cursor are being
  wired up: [Claude Code](/docs/mcp-claude-code),
  [Cursor](/docs/mcp-cursor).

## Learning more

The [protocol reference](/docs/mcp-protocol) covers the JSON-RPC message
shapes; [configuration](/docs/mcp-configuration) covers how a client is
pointed at a server.