Hygraph MCP Server

The Hygraph MCP Server connects your AI assistant directly to a project in the Hygraph headless CMS. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude, Cursor or VS Code with Copilot work with external services. A headless CMS manages content in one place and delivers it to websites and apps through an API. The server comes from Hygraph itself, formerly known as GraphCMS, whose platform has been built on GraphQL from day one. Hygraph hosts the server, so there is nothing for you to run yourself.

What is the Hygraph MCP Server good for?

Plenty of CMS tasks are quick to describe but tedious to click through. Think of finding every draft with a certain tag and changing its priority. Or adding a new field to a content model and then promoting it to the production environment. With the MCP server, you phrase jobs like these in plain language, and the assistant carries them out through the project's API. According to Hygraph, three groups benefit most. Developers can have TypeScript types generated for every content model or export posts as Markdown. Content teams run bulk updates based on criteria. Platform teams handle migrations and clean-ups in conversation instead of writing one-off scripts.

What the server can do

The documentation groups the tools into four areas. On the content side, the assistant reads, filters, creates, updates and publishes entries, and moves content between environments. On the schema side, it inspects and creates content models, adds or changes fields and manages enumerations and components. Then there are Hygraph's AI agents: the assistant can list configured agents, run them on entries and check their status. This only works for agents set to Use in Agent runs, and their output is saved as a draft. A project's AI Guidelines, meaning brand voice, glossary and compliance rules, can be read before the assistant writes anything, but not changed.

Deliberate limits

What stands out is what the server is not allowed to do. Entries and schema elements can never be deleted through MCP, and unpublishing is ruled out as well. Hygraph says this is meant to prevent accidental data loss in automated workflows. You still handle those steps by hand in Hygraph Studio. Anything the assistant creates or changes through the Claude connector starts out as a draft and only goes live once you approve it. One tool still deserves attention: according to Hygraph, execute_graphql can also make bulk changes, so it pays to read the summary before approving each call.

Two ways to sign in

The server offers two endpoints. The project endpoint covers exactly one project and one environment and is called with a permanent auth token sent as a bearer header. The token type decides what the assistant may do: General can edit content and schema, Content only content, Management only the schema, and Read-only can only read. Hygraph recommends separate tokens for staging and production. The global endpoint at mcp.hygraph.com, by contrast, uses your Hygraph account and sees every project you can access. It is mainly used by the official connector in the Claude directory, which needs neither Node.js nor a token.

Clients and setup

Hygraph documents setup for Claude Code, the Claude connector, Claude Desktop, Cursor, VS Code and Windsurf. Claude Code and VS Code talk to the HTTP endpoint directly. Cursor, Windsurf and Claude Desktop go through the mcp-remote helper, which requires Node.js 20.18.1 or later and npx. Cursor also gets a one-click install from Hygraph Studio. One thing to note there: according to Hygraph, Cursor's Auto model mode is not compatible, so you have to pick a model explicitly. Codex is not mentioned in the documentation.

Cost, status and license

The documentation lists all Hygraph projects as eligible and mentions no separate fee. According to its pricing page, Hygraph itself offers a free Hobby plan with no time limit, a paid Growth plan and enterprise tiers sold through its sales team. In its announcement post, Hygraph describes the server as early access, so features may still change. A single FAQ answer in the same post, however, names the enterprise plan as a requirement. It is worth checking under Project Settings and Endpoints whether your own project offers the MCP endpoint. The server has not been released as open source, and there is no public repository.

Who is the Hygraph MCP Server for?

The server suits teams that already work with Hygraph and want to get routine work on content and content models done faster. Developers no longer have to switch between Studio, the API playground and their code, and editorial teams get bulk changes without a script. Compared with the MCP servers of other headless CMSs such as Contentful, Sanity, Strapi or Storyblok, Hygraph stands out for its fine-grained permissions through token types and for leaving out delete rights entirely. If you regularly need to remove or unpublish content, though, you will still do that in Studio.

A Hygraph project with at least one environment and an MCP client such as Claude Code, Claude, Cursor, VS Code or Windsurf. For the project endpoint, a permanent auth token; for Cursor, Windsurf and Claude Desktop, also Node.js 20.18.1 or later and npx.

With the Hygraph MCP Server, your AI assistant works on the content and content models of a Hygraph project. Hygraph hosts the server itself. This guide covers setup in common clients and how to keep permissions sensible.

Step 1: Prepare endpoint and token

For Claude Code, Cursor and VS Code you need the project endpoint. You will find it in Hygraph Studio under Project Settings, Endpoints, MCP Server API. It looks like https://mcp-{REGION}.hygraph.com/{PROJECT_ID}/{ENVIRONMENT}/mcp and covers exactly one environment. Then create a permanent auth token and pick the type that fits the job: General for content and schema, Content for content only, Management for the schema only, Read-only for reading. Store the token in an environment variable such as HYGRAPH_TOKEN.

Claude Code

Run claude mcp add --transport http hygraph <endpoint> --header "Authorization: Bearer $HYGRAPH_TOKEN". Claude Code talks to the HTTP endpoint directly, so mcp-remote is not needed. Use --scope user or --scope project to decide where the entry applies. You can check the connection with claude mcp get hygraph.

Claude

There is an official connector in the Claude directory. Open Settings and Connectors, search for Hygraph MCP and click Connect. Then sign in to Hygraph and confirm with Accept. This route uses the global endpoint and needs neither Node.js nor a token. On Claude Team and Enterprise plans, an owner has to enable the connector first.

Cursor

The easiest option is the one-click install: in Hygraph Studio, click Add to Cursor under Endpoints and MCP Server API. To set it up by hand, add an entry to .cursor/mcp.json with command npx and the arguments -y, mcp-remote@latest, the endpoint, --header and the Authorization value. This requires Node.js 20.18.1 or later. On Windows, use npx.cmd instead of npx. Important: pick a fixed model in Cursor, because according to Hygraph the Auto mode is not compatible.

VS Code and Windsurf

In VS Code, add an entry under servers in .vscode/mcp.json with type http, the endpoint as url and a headers object containing Authorization. Windsurf connects through mcp-remote just like Cursor, using the file ~/.codeium/windsurf/mcp_config.json. Then choose Refresh Servers.

Work safely

Deleting and unpublishing are blocked through MCP anyway. The assistant can still change and publish content and schema, though. Use the narrowest token type that fits and separate tokens for staging and production. The execute_graphql tool can also make bulk changes, so read the summary before approving each call.

claude mcp add --transport http hygraph https://mcp-{REGION}.hygraph.com/{PROJECT_ID}/{ENVIRONMENT}/mcp --header "Authorization: Bearer $HYGRAPH_TOKEN"

Project endpoint: a permanent auth token sent as a bearer value in the Authorization header, with rights set by token type (General, Content, Management, Read-only). Global endpoint and Claude connector: sign in with your own Hygraph account.

Depending on token type or user rights, reading, creating, updating and publishing content plus schema management, as well as running agents and reading AI Guidelines. Deleting and unpublishing are never possible through MCP.

Requests run through the server hosted by Hygraph to the respective project. Content shown in the chat also reaches the client's AI provider. Hygraph recommends passing tokens as environment variables rather than hard-coding them in configuration files.

Medium. The assistant can change and publish content and schema, but cannot delete anything. A crafted text with hidden instructions can trigger unwanted changes, and the execute_graphql tool can make bulk edits. Hygraph recommends fitting token types, separate tokens per environment and reviewing every approval.