# MCP configuration

How an MCP client gets pointed at a server, and what Poly needs from you.

## The shape of a config

An MCP client — Claude Code, Cursor, your own tooling — keeps a list of
servers: a name, a transport (a command or a URL), and credentials.

```json "Typical entry"
{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "some-mcp-server"],
      "env": {
        "MY_API_TOKEN": "the-token"
      }
    }
  }
}
```

Remote servers swap the command for a URL the client connects to. Either
way, the client discovers the server's tools, resources, and prompts at
[initialize](/docs/mcp-protocol).

## On Poly

- **Inside Poly, no config file** — MCP access is a switch. Enable the
  Secrets MCP toggle and agents get vault operations; see
  [Secrets & MCP](/docs/secrets-mcp-integration).
- **External clients** — pointing Claude Code or Cursor at Poly's MCP
  endpoints is still landing. The per-client configs will live on
  [Claude Code](/docs/mcp-claude-code) and [Cursor](/docs/mcp-cursor) when
  they ship.