Supabase MCP in Make: How to let AI agents into your database without allowing everything
Since late October, Make scenarios and Make AI Agents can work with Supabase directly through the MCP Client, on every Make plan. The door is wide open: without limits, an agent can run SQL, apply migrations and deploy Edge Functions. Add the server with project_ref, read_only and features in the URL, and use a fixed tool call instead of letting the AI choose for any routine workflow.
Make announced the new Supabase integration on November 12 in its Community News, after Supabase introduced it on October 28. At its core is the Supabase MCP server. MCP (Model Context Protocol) is an open standard that lets AI agents call ready-made tools from a service, in this case reading tables, running SQL or querying logs. According to Make, it works on every plan, including Free.
What are the ways to connect Make to Supabase?
There are now three, and the main difference is who decides what happens.
| Option | Access | Who decides | Good for |
|---|---|---|---|
| Supabase app with fixed modules | Project ID plus secret key | you, module by module | create a row, search, upsert, upload a file |
| MCP Client, "Call a tool" module | OAuth with your Supabase account | you, one specific tool | a fixed query, e.g. logs or numbers for a report |
| MCP Client "Execute an action with AI" or Make AI Agent | OAuth with your Supabase account | the language model | open questions about your data, sorting errors |
The Supabase app includes modules such as Watch Events, Create a Row, Search Rows, Upsert a Record and Delete Rows. The MCP Client has two modules: "Call a tool" runs exactly one tool with mapped inputs, while "Execute an action with AI" lets a model pick which tools to use for a task you describe. One caveat: Make's own help still lists the MCP Client as an open beta, and pricing may change.
What can an agent do through the Supabase MCP server?
Without limits, a lot. The Supabase docs list these tool groups, among others:
- Database: list tables,
execute_sql(any SQL),apply_migration(change the schema) - Edge Functions: read functions,
deploy_edge_function, create secrets - Debugging: query logs, fetch security and performance advisories
- Account: create, pause and restore projects, check costs
- Branching (experimental, paid plans) and Storage (off by default)
Supabase explicitly warns about prompt injection there: text in your data, such as a customer comment, can try to trick the model into harmful queries. The docs recommend approving every tool call manually, as most MCP clients allow. That is exactly what is missing in a scenario that runs on its own overnight. The Make MCP Client help describes no approval step per tool call, only a choice of which tools the AI can see at all.
Supabase itself calls the built-in confirmation for destructive SQL "not a security guarantee". Make does show before a run whether a tool is marked read-only or destructive, and it logs every run and every tool call. But a log only tells you after the fact what got deleted.
How do you limit access in Make?
With three parameters in the server URL. Supabase documents them as part of the address, for example: https://mcp.supabase.com/mcp?project_ref=abc123&read_only=true.
project_ref=YOUR-PROJECT-IDrestricts the server to one project and turns off the account tools. No agent will accidentally create a paid project or pause yours.read_only=trueruns every query as a Postgres user without write permissions.features=database,docsallows only the tool groups you name. You need this because the docs explicitly mention only database queries, feedback and notebook tools for read-only mode. Whetherdeploy_edge_functionis blocked as well is not stated. If you don't need Edge Functions, simply leave that group out.
In Make, open the MCP Client, add your own server via "+ New MCP Server" and enter the full URL instead of picking Supabase from the list of verified servers. Our take: this is the safer route, because you can see the limits in the URL and they don't depend on a default setting. Then test the connection with "Call a tool" and list_tables before any model gets access.
What does this mean for you?
Most business workflows don't need the AI route at all. The Monday report on new sign-ups that Supabase uses as an example is a fixed query: "Call a tool" with execute_sql on a read-only connection is enough and uses no AI credits. Make bills AI actions through its own provider in credits; our Make vs n8n cost comparison has the details.
The AI route pays off where the question changes every time, for example when an agent should sort new errors from the logs and assign them to a ticket. In that case:
- Never connect your production project with write access. Supabase itself advises connecting production only when the task truly needs production data.
- Write through fixed modules. If a row should be created at the end, let the model only read and decide, then write with a regular module afterward. That keeps writes predictable.
- Treat the secret key like a master key. The fixed Supabase modules in Make use the secret key, and according to Supabase it bypasses every Row Level Security policy. It never belongs in emails, chats or URLs.
- Don't hand it to customers. The MCP server runs with your developer permissions. If your customers should access your app from ChatGPT or Claude, you need your own MCP server, as described in our article on the Supabase MCP server for your own app.
- Get alerted on failures. If an agent comes up empty, you want to know before the report goes missing. Our post on Make Event Subscriptions shows how to set that up in Make.
Our take from projects with Lovable, Supabase and Make: the new integration saves a lot of tinkering with HTTP modules and custom Edge Functions. The biggest mistake would be setting it up with full account access just because that is convenient in the first test. If you want help with that, take a look at our AI automation services.
Frequently asked questions
Does the Supabase integration cost extra in Make?
Make says it is available on all plans, including Free. According to Make's help, the MCP Client is in open beta and pricing may still change; AI actions that run through Make's own AI provider consume credits.
Do I need a Supabase API key for the MCP route?
No. The hosted Supabase MCP server uses OAuth by default: you sign in to Supabase and approve access. Only the fixed modules of the Supabase app need the secret key.
Does Row Level Security protect against a misguided agent?
Only partly. The secret key used by the fixed modules bypasses Row Level Security completely, and the MCP server runs with your developer permissions. Limit access with read_only, project_ref and features instead of relying on RLS.
Does this work with n8n too?
In principle, yes. MCP is an open standard, and the Supabase parameters live in the server URL, not in the tool. The same limits apply there.
Sources
Tell us what you're planning. You'll get an honest assessment within 24 hours.
More articles