# Connect AI tools over MCP

MixShift exposes an MCP (Model Context Protocol) server, so AI tools can query your warehouse, look at live Amazon data, and read your brand context from inside their own chat. Any current MCP client connects over streamable HTTP at `https://mcp.mixshift.io/mcp`.

One distinction worth knowing up front: the [MixShift AI plugin](/knowledge-base/plugin) for Claude bundles skills and a CLI on top of this connection. Every other client below gets the data connection itself, which is already a lot: tables, live Amazon operations, timeline, and brand context as tools the model can call.

## Claude (claude.ai and Claude Desktop)

Add a custom connector pointed at `https://mcp.mixshift.io/mcp` and sign in when prompted. Claude handles the OAuth flow; there is no token to paste. If you use Claude Code or Cowork, skip the connector and [install the plugin](/knowledge-base/plugin/getting-started/install-the-plugin) instead; it is the better experience there.

## ChatGPT

Add MixShift as a connector in ChatGPT's settings, pointed at the same MCP address, and sign in when prompted. ChatGPT registers itself and walks the OAuth flow; like Claude, no token pasting.

## Cursor

Cursor takes a static token. Get one from `mcp.mixshift.io/login?mode=direct`, then add the server to `.cursor/mcp.json` with the token in the `Authorization: Bearer` header. Sessions are long-lived; refresh the token when a request comes back unauthorized.

## Codex

Same pattern as Cursor: token from `mcp.mixshift.io/login?mode=direct`, configured in `~/.codex/config.toml`. Codex requires the streamable HTTP transport, which is exactly what the endpoint speaks.

## Unattended connections

Automations and servers should mint tokens from a service credential rather than a personal login. See [Get access](/knowledge-base/builder-platform/getting-started/get-access).

## What the tools can do once connected

- List and describe warehouse tables, and run read-only SQL.
- Call cataloged live Amazon operations (retail and advertising).
- Read and add to your brand change log, record and corroborate stakes, and read brand context documents. See [Work with brand memory](/knowledge-base/builder-platform/how-to/work-with-brand-memory).

Each tool call checks its own permission scope, so a read-only credential simply cannot invoke a write, no matter what the model asks for.

## Related

- [Get access](/knowledge-base/builder-platform/getting-started/get-access)
- [Limits and guardrails](/knowledge-base/builder-platform/reference/limits-and-guardrails)