Supabase MCP server for your own app: let customers use your web app from ChatGPT and Claude
Since October 22, Supabase ships a ready-made building block that gives your app its own MCP server: customers sign in to ChatGPT or Claude with their usual account and work with their own data there. Your database access rules stay exactly as they are, and extra costs only come from logins and function calls. It pays off for customer portals and internal tools with recurring questions, not for every app.
What did Supabase announce on October 22?
On October 22, Supabase walked through step by step how a Supabase app gets its own MCP server. MCP (Model Context Protocol) is the open standard AI assistants like ChatGPT or Claude use to call outside tools. What is new: it is not you as the developer controlling your project this way. Your customers use your app from inside the chat and sign in with the account they already have.
One command adds the building block to your project (npx shadcn@latest add https://supabase.com/library/r/mcp.json). It creates an Edge Function, a small server function that Supabase runs for you. Inside are the MCP handler, an auth check, and a sample tool called whoami. Every additional tool, such as "show my open tasks", lives in its own file.
Isn't this the same Supabase MCP server that already exists?
No, and this is exactly where many tutorials mix things up. There are two completely different things:
| Supabase's developer MCP server | Your app's MCP server (new) | |
|---|---|---|
| Who uses it? | You and your team in Cursor, Claude Code and similar tools | Your customers or employees in ChatGPT, Claude and other clients |
| What can it do? | Create projects, manage tables and migrations, run SQL | Only the tools you build yourself |
| Permissions | According to Supabase, it can change schema and data | Same permissions the user has in your app |
| Suitable for customers? | No | Yes, built for it |
Our take: if you want to give customers AI access to your app and use the developer server or a server with an admin key for it, you open up your whole database. The new approach is the clean solution for this.
How does each customer's data stay with that customer?
Through the same rules that already protect your app. Every query runs with the identity of the signed-in user, so your Row Level Security policies (RLS, row-level permissions inside the database) apply. Supabase writes that each tool can be a plain query "with no service role key", meaning without the admin key that bypasses RLS.
Sign-in is handled by Supabase Auth acting as an OAuth 2.1 server. OAuth is the familiar flow behind "Sign in with Google": the user confirms on your site that ChatGPT or Claude may act on their behalf. Supabase recommends keeping business rules where your app already checks them and having the tools call that code. That way, an expense approval in the chat follows the same rules as one in your interface.
What steps does the setup take?
The revised guide Deploy MCP servers now covers authentication too. In short:
- Turn on asymmetric signing keys and the OAuth 2.1 server in your Supabase project.
- Enable Dynamic Client Registration so MCP clients can register themselves.
- Build a consent page in your frontend. Supabase only redirects there; you design the page yourself (the "OAuth Consent" block helps).
- Turn off automatic JWT verification for the
mcpfunction (verify_jwt = false), because the function checks the token itself. - Test locally with the MCP Inspector or
claude mcp add, then deploy and give your users the server URL.
Know two limits: Edge Functions are stateless, so MCP sampling (the server asking the AI model for a response) does not work there. And if you do not want Supabase hosting, Supabase says the helpers also run in other backends that use standard web requests, such as Cloudflare Workers.
What does an MCP server for your own app cost?
The OAuth server itself costs nothing extra. Costs come from two meters on the pricing page:
- Monthly active users (MAU): Free includes 50,000, Pro (from $25 a month) and Team (from $599 a month) include 100,000 each, then $0.00325 per user. If one customer uses several agents, Supabase still counts them only once.
- Edge Function invocations: Free includes 500,000, Pro and Team include 2 million each, then $2 per million. Every tool call from the chat triggers at least one function invocation.
A typical customer portal with a few hundred users stays within the included amounts on the Pro plan. The Free plan is only good for trying things out: projects there are paused after one week without activity.
What does this mean for you?
An MCP server is worth it when your users regularly ask questions about their data: "Which orders are open?", "How many leads came in this week?", "Create the invoice for customer X". Typical cases are customer portals, B2B software, and internal tools that currently replace a spreadsheet. For a pure marketing website or an app customers open once a month, the effort is too high.
Our checklist before you start:
- Check RLS: Is every table a tool reads really protected by row-level permissions? Without RLS, the chat sees more than your interface does.
- Read first, write later: Start with two to three read-only tools. Add tools that write data only where your app already enforces the rule on the server.
- Lock down registration: Supabase itself warns that with dynamic registration any MCP client can register. So always require explicit user approval, check redirect URLs, and review registered clients regularly.
- Test your clients: Through the MCP Apps extension, tools can also show interfaces in the chat, such as cards or forms. Support varies by client. The list there includes Claude, VS Code with GitHub Copilot, and Microsoft 365 Copilot, and Supabase shows its demo in ChatGPT.
If you build with Lovable: since October 20, according to its changelog, Lovable can create your own Supabase project when the built-in backend is turned off. The region cannot be changed afterward, and Supabase bills your own organization. Where your data lives in that case is covered in our article on Lovable and GDPR. How an idea becomes a solid product is in our guide to MVP development.
Our take: for web apps, this is the first official way to offer ChatGPT and Claude as a second interface without maintaining permissions twice. The real work is not the building block but picking the right three tools. If you want help with that, take a look at our web app services or reach out to us.
Frequently asked questions
Do my customers need a paid ChatGPT or Claude plan?
That is up to each provider, not Supabase. Before you launch, check with OpenAI and Anthropic which plans let users connect custom MCP servers, and test with the clients your customers actually use.
Can the AI see other customers' data through the MCP server?
Not if every affected table is properly protected with Row Level Security and no tool uses the service role key. Then the chat follows exactly the same row-level permissions as your app.
Does this work with an app built on Lovable Cloud?
The building block needs an Edge Function and the OAuth 2.1 server in a Supabase project you fully control. With your own Supabase project, that is the direct route. With a managed backend, check first whether you can change these settings.
What happens when a user wants to revoke access?
Supabase recommends letting users deny an MCP client's access and revoke it later. Build this into your account settings from day one.
Sources
Tell us what you're planning. You'll get an honest assessment within 24 hours.
More articles