Reference build — GPTs as your agent

Turn a Custom GPT into a real agent.

One Convex backend, one set of Actions. A Custom GPT, an agent harness (Claude Code), or the dashboard — all three write into the same workspace.

Demo data and your own data live in separate tenants.

Three clients, one backend

/api/agent/*token-scoped

Convex deployment

One backend, three clients

A pattern you can lift.

Start on the shared demo tenant, then deploy your own and point your GPT at it.

The GPT, Claude Code, and the dashboard all call the same /api/agent/* routes. No client-specific backend.

One token resolves to exactly one business id, so it can never read or write another tenant's rows.

Reads and writes land in the activity log with the action name, so agent behaviour stays auditable.

Read and write Actions share one HTTPS surfaceThe backend runs headless — the dashboard is optionalTokens are stored hashed and shown once