Add your database to Claude with one URL

Hosted AI clients have a bootstrapping problem: there's nowhere to paste an API key. When you add a connector in Claude or ChatGPT, all you get to type is a URL. Everything else — what this service is, how to authenticate, what's allowed — the agent has to discover on its own.

As of this week, SchemaStack speaks that language end to end. Give the connector this, with your workspace's UUID:

https://schemastack.io/mcp?workspace=<workspaceUuid>

The UUID is in the address bar when the workspace is open in the admin app. That's the entire setup.

What actually happens

The first request arrives with no credential, and instead of a dead end it gets a 401 pointing at a discovery document:

HTTP/1.1 401 Unauthorized
WWW-Authenticate: Bearer realm="SchemaStack MCP",
  resource_metadata="https://schemastack.io/.well-known/oauth-protected-resource/mcp"

That document names the authorization server; the server's own metadata lists the endpoints; the client registers itself and gets a client_id. Registering grants nothing — a registered application has reached no data yet. It has a name and a place to send you.

Then your browser opens the consent screen. It names the application, the workspace, and exactly what's being asked. One thing we made sure it says out loud: when an application registered itself rather than being added by an administrator, the screen tells you so — because anyone can register a client under any name, and the name is a claim, not an endorsement. Only approve connections you started yourself.

Approve, and the agent receives a token for that one workspace. It can now use the same tools our own interface uses — around 57 of them, from querying rows to previewing schema migrations.

Scoped below your schema

Consent uses two scopes, and they cap what any token can ever do:

Notice what's missing: no scope reaches your schema. Agreeing to "write" means letting an application write rows — not letting it drop a column, change a type, or re-import your schema on a database you own. Structural access exists only through mcp_ API keys, which an administrator creates deliberately, and even those are capped by a per-workspace access level that starts at disabled.

Revocation that means it

Access tokens are stateless, which usually means revoking one is a polite request that takes effect when the token expires. We didn't think "your agent loses access within the hour" was an acceptable sentence, so revocation here records the moment it happened — and any token issued before that moment is refused from then on, on MCP and on the REST API alike.

Workspace → OAuth2 lists every application with an active grant, including self-registered ones, and who approved them. Revoke the application or one person's grant; either way it bites immediately.

Prefer keys? They still work

Claude Desktop, Claude Code and Cursor read config files, so they can keep using mcp_ keys — which do things OAuth tokens deliberately can't: carry a higher access level (up to schema work) and scope down to named views.

{
  "mcpServers": {
    "schemastack": {
      "url": "https://schemastack.io/mcp",
      "headers": { "Authorization": "Bearer mcp_your_key_here" }
    }
  }
}

What it doesn't do (yet)

• An OAuth token covers one workspace — connecting a second workspace is a second connector.
• OAuth tokens can't be scoped to individual views; that's mcp_ keys only.
• There's no organisation-wide consent policy ("only admins may approve connectors") — any signed-in member with workspace access can grant it today.
• ChatGPT's connector support depends on your ChatGPT plan; Claude's works on claude.ai and Claude Desktop.

Everything above is live. The AI integration guide has the full walkthrough, including the discovery documents if you're building your own client.