Skip to main content
The studio runs its own Model Context Protocol (MCP) server, so Claude or any other MCP client can read and update the business: proposals, agreements, invoices, design reviews, activity, the photo library and brand kits. This page covers the endpoint, how to authenticate, how the server behaves, and every tool it exposes.

How the server works

The server lives at /api/mcp/ on each deployment. It speaks the Streamable HTTP transport as stateless JSON-RPC 2.0 and answers POST only. There are no sessions and no server-initiated stream, so an authenticated GET or DELETE returns 405. It supports four methods: initialize, ping, tools/list and tools/call. A request can be a single message or a batch, and a batch is answered in the order it was sent. Notifications get no response; a request made only of notifications returns 202. During initialize the server negotiates one of the protocol versions 2025-06-18, 2025-03-26 or 2024-11-05, falling back to the newest when the client asks for another.

Authenticate

Access is controlled by one environment variable, MCP_ACCESS_TOKEN. It must be at least 32 characters long. If it is missing or shorter, every request is refused with 401, so the endpoint is effectively off.
Claude’s custom connector takes a URL only, so the token goes in the path:
The token is compared in constant time.
The URL form is a password. Anyone with the full address can read proposals and invoices. To revoke every copy at once, rotate MCP_ACCESS_TOKEN in Vercel and redeploy.

Connect Claude

1

Set the token

Set MCP_ACCESS_TOKEN on the deployment to a value of 32 characters or more, then redeploy so the function reads it.
2

Add the connector

In Claude, add a custom connector with the URL form above, using your site’s domain and the token.
3

Check the tools

Claude calls tools/list and shows the tools below. Tools that change data are annotated as not read-only, so Claude can ask before calling them.

Site tools

draft_invoice only saves a draft. You review and send it from Admin → Invoices, which creates it in Stripe and emails the client a pay link. See Proposals, contracts and invoices.

Brand tools

Every brand tool except list_brand_kits takes kit, which accepts a brand kit’s id or its name (an exact match first, then a unique prefix, ignoring case). Each write saves a new version of the kit, so earlier versions stay in its history.
A kit holds at most 60 references. When add_brand_reference reports the kit is full, remove some with remove_brand_reference first. For more on kits, see Brand kits.

Errors

A problem you can fix, such as a bad input or a kit that cannot be found, comes back as a tool result with isError set and a message Claude can read. Any other failure returns “The site couldn’t do that just now.” An unknown tool or method is a JSON-RPC error.

Calling the tools from scripts

scripts/siteConnector.mjs calls these tools from local scripts, including the DESIGN.md sync and the website scan. It reads MCP_ACCESS_TOKEN from .env.local, and DESIGN_MD_SITE points it at another deployment. The repo’s own .mcp.json is unrelated: it gives local coding agents a chrome-devtools MCP server. For environment setup, see Local setup.