# Overview (/docs/api/mcp-server/overview)



The **Layers MCP server** exposes the Partner API as [Model Context Protocol](https://modelcontextprotocol.io) tools. An MCP client (Claude, ChatGPT via the OpenAI Agents SDK, Cursor, or any other) spawns it, and the agent can then create projects, generate content, schedule posts, and read metrics by calling tools directly — without you writing a single `fetch`.

It is published to npm as [`@layers/mcp-server`](https://www.npmjs.com/package/@layers/mcp-server) and runs locally over stdio. There is no hosted endpoint to point at and no service to operate.

<Callout type="info">
  **It's a thin wrapper, by design.** Every tool maps to exactly one documented
  `/v1` endpoint. The MCP server adds no new capabilities — it's an ergonomic
  surface over the same API, so it tracks the [API reference](/docs/api/reference)
  one-to-one. When you want the precise request/response shape for a tool, read
  its linked endpoint page.
</Callout>

## Who it's for [#who-its-for]

* **Agent builders** wiring Layers into an AI assistant or autonomous workflow, who want tool-calling instead of an HTTP client.
* **Anyone prototyping** against the Partner API who'd rather say "list my Layers projects" in an MCP client than hand-roll requests.

If you're integrating Layers into a conventional backend, you don't need the MCP server — call the [Partner API](/docs/api/getting-started) directly.

## How it relates to the Partner API [#how-it-relates-to-the-partner-api]

The MCP server is a client of the Partner API. It authenticates with the same `lp_...` key, talks to the same `https://api.layers.com/v1` host, and inherits the same auth, scopes, idempotency, rate limits, and sandbox behavior.

That means everything you already know about the API carries over:

* **Auth & scopes** — the key you pass the server determines what it can do. See [Authentication](/docs/api/getting-started/authentication).
* **Sandbox** — a `lp_test_...` key returns fixture-backed results, so an agent can be developed safely against it. See [Sandbox](/docs/api/concepts/sandbox).
* **Async jobs, pagination, idempotency, errors** — identical semantics, surfaced through MCP conventions. See [How it works](/docs/api/mcp-server/how-it-works).

## Next steps [#next-steps]

<Cards>
  <Card title="Quickstart" href="/docs/api/mcp-server/quickstart" description="Connect the server to Claude or the OpenAI Agents SDK in under a minute." />

  <Card title="Authentication & configuration" href="/docs/api/mcp-server/authentication" description="Flags and env vars: API key, base URL, read-only mode, child-org acting." />

  <Card title="How it works" href="/docs/api/mcp-server/how-it-works" description="Transport, server instructions, tool annotations, async jobs, pagination, idempotency, errors." />

  <Card title="Tool reference" href="/docs/api/mcp-server/tool-reference" description="All 52 tools, grouped by domain, each linked to its API reference page." />
</Cards>

## Package [#package]

* **npm:** [`@layers/mcp-server`](https://www.npmjs.com/package/@layers/mcp-server)
* **Source:** [github.com/layers/mcp-server](https://github.com/layers/mcp-server)
* **License:** Apache-2.0