Supabase + ChatGPT
Connect Supabase to ChatGPT
with OpenAI models and the Agent200 SDK
ChatGPT cannot send an API key to a custom MCP server, so this guide builds Supabase into your own agent with OpenAI's models and the Agent200 SDK.
On this page
Overview
ChatGPT can add custom MCP servers, but it signs in to them only with OAuth or with no sign-in, and cannot send an API key. Agent200’s MCP server uses your API key, so this guide takes the SDK route: you build Supabase into your own agent with OpenAI’s models, through Agent200.
Your agent then calls the Supabase tools with each user’s own Supabase connection: Agent200 runs the OAuth flow, stores the tokens and keeps them fresh.
- Connector
- Supabase
- Works with
- ChatGPT · Your own agent code
- Authentication
- Managed OAuth, per end user
- You need
- An Agent200 API key
- Tools
- 6 tools
- Model in examples
openai/gpt-5.6-sol
Before you start
- An Agent200 API key. Create one on the API keys page of the dashboard.
- Supabase enabled for your project. In the dashboard's Connectors settings, with the endpoints and scopes your agent needs.
- An end user with a Supabase account. They approve access once, on Supabase's own page.
- A model for your agent. The examples use openai/gpt-5.6-sol, through the same Agent200 API key.
Connect Supabase to ChatGPT
-
1
Step 1: Create your Agent200 API key
Open the API keys page in the Agent200 dashboard and choose Create an API key. Copy the key and store it as an environment variable, AGENT200_API_KEY, or in your secrets manager. One key covers Supabase and every other connector and model enabled for your project. Keep it out of chats, prompts and source control.
.envConceptual exampleStore the Agent200 API key as an environment variable # Keep the key out of your code, prompts and chats. export AGENT200_API_KEY="paste-your-key-here" -
2
Step 2: Set up the Agent200 SDK
Add the SDK to your project and create a client with the key from the environment.
agent.jsConceptual exampleSet up the SDK with the Agent200 API key // One Agent200 API key covers every connector enabled for the project. const agent200 = new Agent200({ apiKey: process.env.AGENT200_API_KEY, }); -
3
Step 3: Let your user connect Supabase
Each end user authorizes their own Supabase account once, through managed OAuth. If the user has not connected yet, Agent200 returns an authorization link: show it to them, and they approve access on Supabase's own page. Agent200 handles the callback, stores the tokens and refreshes them.
connect.jsConceptual exampleLet an end user connect their account (managed OAuth) // Each end user authorizes their own Supabase account, once. const connection = await agent200 .user("user_123") .connect("supabase"); if (connection.authorizationUrl) { // Show this link to your user; Agent200 handles the rest. console.log(connection.authorizationUrl); } -
4
Step 4: Run with the Supabase tools
Name the user, the model and Supabase in the tools. The model decides when a Supabase tool is needed, and Agent200 executes the call with the user's own connection.
run.jsConceptual exampleRun a request with the connector's tools const result = await agent200 .user("user_123") .run({ model: "openai/gpt-5.6-sol", input: "How many users signed up yesterday", tools: ["supabase"], }); console.log(result.output);
Test your connection
Run one request to check everything is wired up.
Ask
How many users signed up yesterday
What should come back
-
An answer in
result.output, built from the user's own Supabase data. -
Behind the scenes, the model calls a Supabase tool, such as
supabase_list_projects, and Agent200 runs it.
If Agent200 returns an authorization link instead, the user has not connected Supabase yet. Go back to the connect step.
Get your API key
Create a key in the Agent200 dashboard. It reaches Supabase and every other connector enabled for your project.
Instructions for AI agents
Using a coding agent? Paste this brief into it, or give it the address of this page. It lists what to do, in order, to connect Supabase to ChatGPT.
Goal: connect Supabase to ChatGPT with Agent200. Guide: https://www.agent200.com/tools/supabase/framework/chatgpt/ 1. Get an Agent200 API key. If the developer has none, send them to https://agent200.com/platform/apikeys/ and ask them to choose "Create an API key". Read the key from the AGENT200_API_KEY environment variable. Never ask for it in chat, and never write it into code, prompts or logs. 2. Make sure Supabase is enabled for the project in the Agent200 dashboard (Connectors), with the endpoints and scopes the task needs. 3. Add the Agent200 SDK to the project and create a client with that key. 4. Give every request a stable id for the end user. Agent200 uses only that user's own connections. 5. Connect the user's Supabase account through Agent200. If Agent200 returns an authorization link, show it to the user and continue once they have approved access on Supabase. 6. Run requests through Agent200 with a model (for example openai/gpt-5.6-sol) and "supabase" in the tools. The model decides when a Supabase tool is needed; Agent200 executes the call. 7. Check it works with this request: "How many users signed up yesterday" What should come back: - An answer in `result.output`, built from the user's own Supabase data. - Behind the scenes, the model calls a Supabase tool, such as `supabase_list_projects`, and Agent200 runs it. 8. If a tool is missing or a call is refused, check the project's Connectors settings in the dashboard and the access the user granted in Supabase, then retry. The code on the guide page is a conceptual example of the SDK. SDK reference: https://docs.igpt.ai/
The whole guide is also available as Markdown, for agents and LLMs: /tools/supabase/framework/chatgpt.md
Supabase tools ChatGPT can call
-
List projects
Readsupabase_list_projectsList the projects the connected user can access.
-
Run a read-only query
Readsupabase_run_read_only_queryRun a SQL query that can read but never change the database.
-
Run a SQL query
Writesupabase_run_queryRun a SQL statement, including ones that change data or schema.
-
List edge functions
Readsupabase_list_functionsList a project's edge functions and their status.
-
Get logs
Readsupabase_get_logsRead a project's logs for a time range.
-
Generate TypeScript types
Readsupabase_get_typescript_typesGenerate TypeScript types from the database schema.
No tool matches that search.
How it works
-
01
Your app sends the request
It calls Agent200 with the model, instructions, the input, the end user's id and the Supabase tools.
-
02
The model picks a tool
The model decides whether a Supabase tool is needed, and with which arguments.
-
03
Agent200 runs the call
It uses that user's own Supabase connection, within the endpoints and scopes your project allows, and returns the result to the model.
-
04
Your app gets the result
Agent200 returns the completed result. Your app decides when the agent runs and what happens next.
Each user's own connection. Connections belong to individual end users. A request for one user never reaches another user's Supabase account.
Permissions at every level. Your project settings decide what the agent may request, the end user decides what they authorize, and Supabase enforces what its API grants.
Troubleshooting
Why not add Agent200 to ChatGPT as an MCP server?
ChatGPT’s custom MCP servers sign in with OAuth or with no sign-in, and cannot send an API key. Agent200’s MCP server uses your API key, so for ChatGPT users the SDK route on this page is the one that works.
The request says Supabase authorization is required
The end user has not connected Supabase yet, or the connection needs to be approved again. Agent200 returns an authorization link: show it to the user, let them approve access on Supabase's own page, then run the request again.
ChatGPT does not use the Supabase tools
Check that Supabase is enabled for your project in the dashboard's Connectors settings and that it is in the tools you send with the request. Naming Supabase in the request, or in the instructions, also helps the model choose the right tool.
A tool call is refused or returns less than expected
Agent200 runs only what your project allows (services, endpoints and scopes), and Supabase returns only what the end user authorized. Widen the endpoints or scopes in the dashboard if the agent needs more, then ask the user to connect Supabase again so they can approve the new access.
The results come from the wrong account
Every request names an end user (user_123 in the examples). Pass the id of the person your agent is working for: Agent200 then uses only that user's Supabase connection, never another user's.
The API key was pasted into a chat or committed
Create a new key on the API keys page, update AGENT200_API_KEY wherever your app reads it, and stop using the old one. Keep keys in environment variables or a secrets manager, never in prompts, chats or source control.