Hyperstack MCP Connector

Hyperstack MCP Connector

Overview

Expose Convex functions as tools that Claude and ChatGPT can call as the signed-in user, over a hosted remote MCP server with real OAuth. One server works across every MCP client. The whole thing lives inside Convex — no separate web server, no bridge, no keypair.

Built on the convex-mcp-gateway component. It hosts the MCP server as a Convex httpAction, handles the JSON-RPC/Streamable-HTTP protocol, the OAuth discovery doc, and an audit log. You supply three things:

1. Tools — defineMcpQuery / defineMcpMutation / defineMcpAction map a Convex function to a tool. Each identity-scoped tool takes an injected caller arg (identityArg), because Convex strips ctx.auth across the component boundary — so you scope on caller.subject, not ctx.auth.
2. Identity — a resolveIdentity that validates Clerk's opaque OAuth access token at Clerk's OIDC userinfo endpoint and returns { subject, claims }. Clerk is the OAuth authorization server (with Dynamic Client Registration); the gateway is the resource server.
3. Mount — one httpAction in convex/http.ts that calls gateway.handleMcpRequest(...), plus the protected-resource discovery route.

Claude/ChatGPT --OAuth (DCR)--> Clerk (authorization server)
      |  opaque access token
      v
convex.site/mcp  (gateway httpAction, requireAuth)
      |  resolveIdentity → Clerk userinfo → { subject }
      v
gateway injects `caller` → internal Convex fn scopes by caller.subject
      (runs as that user; every owner-scoped function works unchanged)

When to use

• You want an AI client to do something in your app (save, fetch, recall, trigger) and have it run as the actual user.
• Your backend is Convex + Clerk — Convex is your source of truth.
• You want it addable in Claude and ChatGPT without per-client backends.

Not for: local stdio MCP servers (no OAuth needed — a token is enough); apps not on Convex+Clerk (the identity resolver assumes Clerk; the tool model assumes Convex components); read-only public data with no user identity (skip OAuth entirely).

Why not host the MCP server in a Next.js/Vercel route and bridge to Convex? You can, but for a Convex-centric app it adds a separate server, a signed bridge JWT, a JWKS endpoint, and key rotation — all to reproduce an identity Convex can already carry. The gateway keeps everything in convex/.

Build steps

1. Install + register: npm install convex-mcp-gateway, then app.use(mcpGateway) in convex/convex.config.ts (setup-and-gotchas.md §1).
2. Tools — copy references/mcp.ts to convex/mcp.ts. Replace the sample tools with one defineMcp{Query,Mutation,Action} per function you expose. Every identity-scoped tool declares caller: mcpCallerValidator + identityArg: "caller" and points fn at an internal function. Keep authorize, resolveClerkIdentity, and initializeInstructions.
3. Tool impls — copy references/mcp-tools.ts to convex/mcpTools.ts. These are internalQuery/internalMutation/internalActions that owner-scope by caller.subject (map it to your user via whatever index keys users on the Clerk id).
4. Mount — merge references/http.ts into convex/http.ts (the /mcp routes + protected-resource discovery + requireAuth + exposed headers).
5. Env — set MCP_AUTH_SERVER_URL (your Clerk issuer) per deployment, or rely on an existing CLERK_JWT_ISSUER_DOMAIN (setup-and-gotchas.md §2).
6. Enable Dynamic Client Registration in the Clerk dashboard — required, one manual toggle; leave "access tokens as JWTs" OFF (setup-and-gotchas.md §3).
7. Deploy (npx convex deploy), then add the connector at https://<deployment>.convex.site/mcp in Claude/ChatGPT (setup-and-gotchas.md §4).

Reference files

Common mistakes

• Reading the user from ctx.auth in a tool → always null. The gateway strips ctx.auth across the component boundary; use the injected caller arg.
• Exposing tool functions as public (api.*) → any Convex client can call them with a forged caller and read another user's data. Make them internal.
• Skipping Dynamic Client Registration in Clerk (or leaving it unsaved) → the client can't register, OAuth never starts. registration_endpoint must appear in Clerk's oauth-authorization-server metadata (not openid-configuration).
• Omitting requireAuth: true → browser clients (claude.ai) see an empty tools/list, think they're connected, and never start OAuth (they only react to a 401).
• Not exposing mcp-session-id / www-authenticate on CORS → the browser client can't read the session or the auth challenge.
• Convex subject mismatch → tools authenticate but owner-scoped queries return nothing. caller.subject (the Clerk user id) must be what your functions key on.
• Weak tool descriptions → the model defaults to web search instead of your connector. Name the concrete data and say when to use it; set initializeInstructions. Re-sync by starting a new chat after changing them.

See references/setup-and-gotchas.md for the full gotcha list (opaque-vs-JWT tokens, the circular-type annotation, ChatGPT caveats, DCR verification) and the round-trip test.

Reference implementation

Padscanner's connector (github.com/tfohlmeister/convex-mcp-gateway is the component; padscanner exposes its rental-search tools to Claude as the signed-in user) is built exactly this way and runs in production — the same pattern powers a savethis "save & recall from Claude" connector. Use it as the worked example, but keep this skill stack-generic.