Base44
Build and edit Base44 apps from an AI assistant, and query data your app's users submit.
Run SQL, apply migrations, deploy Edge Functions and manage projects from an AI assistant.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The Supabase MCP server connects a Supabase account to Claude, ChatGPT and any MCP-compatible agent. It exposes 32 tools for running SQL, applying migrations, reading logs, deploying Edge Functions, managing storage config and creating projects and branches. Sign-in is OAuth, and three URL parameters narrow what the connection can reach.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
secrets:read is a grantable scope, and its platform reach is wider than the connector's tools. Supabase's Management API specification attaches secrets:read to seven read endpoints covering project secrets, API keys and JWT signing keys, and the secrets response schema has a required value field. The only tool that touches any of them is get_publishable_keys, which hardcodes reveal to false and returns client-safe keys only. Grant deliberately: this is the one scope on the consent screen whose name describes credentials.execute_sql runs arbitrary SQL. There is no statement allowlist. Its reach is whatever the connected Postgres role permits, DDL included, unless you connect with read_only=true.read_only=true if you must use real data.features=storage.plain PKCE is still accepted. The authorization server advertises both S256 and plain in code_challenge_methods_supported. Most servers in this catalogue advertise S256 alone.features, read_only and project_ref each change which tools are registered, so no single count describes every connection. 32 is the full catalogue.query_logs appears in Supabase's docs; the directory and the published source both have get_logs.Only publishable keys through its tools. The get_publishable_keys tool calls Supabase's API-keys endpoint with reveal set to false and filters the response to anon and publishable keys, per the server's published source. The secrets:read scope itself is broader, so grant it only when you accept that reach.
Append read_only=true to the endpoint URL. Supabase documents this parameter as executing all queries as a read-only Postgres user. In the published source the flag also hides every write tool from the tool list and makes apply_migration throw before it runs, so the restriction is enforced twice.
Supabase advises against it. Its security documentation tells you not to connect the server to production and to use a development project with non-production or obfuscated data instead. The same page states the server is designed for development and testing purposes only.
It wraps SQL results in an untrusted-data boundary. The published source surrounds every execute_sql result with a randomly generated delimiter and instructions telling the model never to follow commands inside it. Supabase calls this not foolproof and asks you to keep manual tool-call approval enabled.
Yes, but only after a confirmation step it cannot skip. Both create_project and create_branch require a confirm_cost_id, and the published source rejects the call unless that value hashes to the current cost. Supabase's pricing lists branches at 0.01344 dollars per branch per hour.
Storage is the only group off by default. Supabase's MCP documentation states all feature groups except Storage are enabled by default, and the published source confirms storage is absent from the default feature list. Add features=storage to the endpoint URL to enable list_storage_buckets, get_storage_config and update_storage_config.
A paid plan. Supabase's MCP documentation marks the branching group experimental and notes it requires a paid plan, and its pricing page lists branching as not included on the Free Plan and available from the Pro Plan upward at 0.01344 dollars per branch per hour.
https://supabase.com/docs/guides/ai-tools/mcp, read as Markdown at https://supabase.com/docs/guides/ai-tools/mcp.md via the .md suffix the host advertises in llms.txt) · retrieved 2026-08-18github.com/supabase/mcp (cloned and read 2026-08-18; repository linked from the MCP documentation page). Tool registrations, annotations, feature-group defaults, read-only enforcement, the untrusted-data boundary, the cost-confirmation hash check and the reveal: 'false' constant were all read from packages/mcp-server-supabase/src. · retrieved 2026-08-18secrets:read endpoint set, the SecretResponse and ApiKeyResponse schemas, and the reveal parameter. · retrieved 2026-08-18resource_documentation field of the server's own OAuth descriptor, and served over unauthenticated GET from the same host that returns 401 to POST) · retrieved 2026-08-18supabase.com/robots.txt — Allow: / with Content-Signal: ai-train=yes, search=yes, ai-input=yes (retrieved 2026-08-18). Supabase expressly permits AI synthesis of this documentation. · retrieved 2026-08-18https://mcp.supabase.com/mcp — anonymous initialize returned HTTP 401 with WWW-Authenticate: Bearer error="invalid_request" and a resource_metadata pointer; /.well-known/oauth-protected-resource/mcp returned HTTP 200 with 13 scopes and authorization server https://api.supabase.com; api.supabase.com/.well-known/oauth-authorization-server returned HTTP 200 with the same 13 (2026-08-18) · retrieved 2026-08-18Connect Supabase once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.