Contentstack MCP Server

The Contentstack MCP Server connects your AI assistant to Contentstack, a headless CMS built for companies. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude, Cursor, Codex or ChatGPT 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 Contentstack itself. It is available in two forms: the hosted MCP Profile Hub and a local npm package. Contentstack now recommends the hosted route, and its developer site says plainly that Profile Hub replaces the local package.

What is the Contentstack MCP Server good for?

A large CMS involves a lot of clicking. Adding a new field to several content types, collecting entries for a release, checking a translation or finding out why a deployment failed: you describe jobs like these to the assistant in plain language, and it carries them out through Contentstack's APIs. Developers no longer have to jump between code, documentation and the web interface. Content architects can draft and adjust models in conversation. Editorial teams can have entries summarized or drafts written to a brief.

MCP Profile Hub: the hosted route

Profile Hub is an app from the Contentstack Marketplace. In it, you create profiles. A profile is its own MCP connection that shows only the tools you pick for it. Once you publish it, you get a connector URL to paste into your client. On first connect, you sign in to Contentstack with OAuth in the browser. OAuth is a sign-in method where you approve access without handing over a password or key. That is why the URL itself carries no secrets.

Every call runs with the rights of your existing Contentstack role. According to Contentstack, a read-only profile cannot publish anything. Each tool call is recorded under Executions, including the profile, tool, client, timing, input and result. Executions only names the client, though, not the person. For per-user attribution, Contentstack points to the organization audit log. Another practical touch is that the tool descriptions include the real content types, locales, branches and environments of your stack. The assistant guesses less and is more likely to pick valid values.

The local npm package

The @contentstack/mcp package runs on your machine through npx and talks to the client over stdio, meaning standard input and output. According to the documentation, it covers ten tool groups: content management with entries, assets, content types, branches, releases and taxonomies, plus content delivery for published content, analytics, automations, Brand Kit, Launch, Developer Hub, Lytics and Personalize. Only the cma group is active by default. You sign in once with npx @contentstack/mcp --auth or use a management token. The npm registry lists the package under the MIT license, but the repository linked there is not publicly reachable at the moment.

Limits and risks

Both routes can write, and some tools in the local package delete things permanently, such as environments, global fields or asset folders. So pick only the tools or groups you actually need. According to Contentstack, this also makes the model more accurate, since it has fewer tools to choose from. In Profile Hub, injected lists are capped at 50 values, and because of a cache, new content types can take up to five minutes to show up.

Clients

Contentstack names Claude, Claude Code, Cursor, Codex and ChatGPT. It says Claude Code is the most reliable, while Claude Desktop can connect intermittently. On Claude.ai, use depends on your organization's connector settings, and with Codex you should pick a recent model. Clients without HTTP support can reach Profile Hub through the mcp-remote helper, which needs Node.js.

Cost and license

Contentstack does not mention a separate fee for MCP access. According to its pricing page, it offers a free plan with no expiry date, paid plans and enterprise tiers. Agent OS is part of every plan. Profile Hub is a hosted service without published source code, while the local package is MIT-licensed.

Who is the Contentstack MCP Server for?

It is mainly worth it for teams that already run Contentstack in production and want to get model maintenance, releases or Launch deployments done faster. Compared with the MCP servers of other headless CMSs such as Contentful, Sanity, Storyblok or Hygraph, the freely assembled profiles and the log of every call stand out. If you need to limit and trace access closely, you will find good tools here.

A Contentstack account with access to a stack and an MCP client such as Claude Code, Claude, Cursor, Codex or ChatGPT. For Profile Hub, the installed Marketplace app and a published profile; for the local package, Node.js 18 or later and the stack API key.

With the Contentstack MCP Server, your AI assistant works directly with entries, content types and releases in Contentstack. Contentstack recommends the hosted MCP Profile Hub for this. This guide covers setup and the local route as an alternative.

Step 1: Install Profile Hub

Open the Marketplace in Contentstack and install the MCP Profile Hub app. This is where you manage profiles. A profile is its own MCP connection with exactly the tools you allow. If one of the ready-made system profiles fits, you can use it right away.

Step 2: Create your own profile

Create a new profile and choose its tools. Take only what the job needs: fewer tools make the model choose more accurately and save context. Then configure the stack and publish the profile. You get a connector URL in the form https://mcp-profile-hub.contentstack.com/api/mcp?profile_id=<PROFILE_ID>&org_uid=<ORG_UID>. It contains no secrets.

Claude Code

Run claude mcp add --transport http contentstack "<connector URL>". Put the URL in quotes, since it contains an ampersand. On the first call, a browser opens, you sign in to Contentstack and approve the requested permissions. According to Contentstack, Claude Code is the most reliable of all clients.

Cursor, Claude, Codex and ChatGPT

Cursor and Claude have Add to Cursor and Add to Claude buttons. Alternatively, add the URL as a remote server. In Claude.ai, use depends on your organization's connector settings. In Codex, pick a recent model. Clients that only speak stdio can reach the URL through the mcp-remote helper, which needs Node.js.

Local route with the npm package

If you prefer to work locally, use the @contentstack/mcp package. Sign in once with npx @contentstack/mcp --auth. Then add the command npx -y @contentstack/mcp to your client and set CONTENTSTACK_API_KEY and GROUPS as environment variables, for example GROUPS=cma. For servers without a browser, there is the management token, in which case you have to set the region explicitly. Contentstack names Profile Hub as the successor to this package.

Work safely

Every call runs with the rights of your Contentstack role. For analysis, a read-only profile that cannot publish is a good fit. Check the Executions area regularly, where every tool call is listed with its input and result.

claude mcp add --transport http contentstack "https://mcp-profile-hub.contentstack.com/api/mcp?profile_id=<PROFILE_ID>&org_uid=<ORG_UID>"

Profile Hub: OAuth in the browser on first connect, and the URL contains no secrets. Local package: OAuth through npx @contentstack/mcp --auth or a management token for content management; content delivery uses a delivery token, and Lytics its own access token.

Depending on the profile or tool group, reading, creating, updating and publishing entries and assets, managing content types, branches, releases and environments, plus Launch, Personalize and analytics features. Every call runs with the rights of your own Contentstack role, and a read-only profile cannot publish.

With Profile Hub, requests run through the service hosted by Contentstack, and every call is logged with input and result under Executions. The local package stores OAuth tokens on your own machine. Content shown in the chat also reaches the client's AI provider.

Medium to high, depending on what you enable. The assistant can change and publish content, and some tools in the local package delete permanently. A crafted text with hidden instructions can trigger unwanted actions. Enable only the tools you need, use read-only profiles for analysis and review Executions regularly.