Supabase documents self-hosting and MCP

Supabase has published a fuller map for two jobs developers keep asking about: running Supabase themselves with Docker and wiring AI clients into Supabase through Model Context Protocol. (supabase.com 1) (supabase.com 2) The self-hosting guide says Docker is the fastest and recommended path, starting from a cloned Supabase repository and a Docker Compose stack. It lists 4 gigabytes of RAM as the minimum and 8 gigabytes or more as the recommended baseline. (supabase.com 1) (supabase.com 2) Supabase also says operators can trim the stack by removing services they do not need, including Logflare analytics, Realtime, Storage, imgproxy, and Edge Runtime. That turns the guide into a menu for smaller development or small-to-medium production setups, not just a one-size install. (supabase.com) Model Context Protocol, or MCP, is the interface that lets an AI client call outside tools instead of only generating text. Supabase’s hosted MCP server is aimed at that use case, letting a client connect to a Supabase project and query tools after the user signs in. (supabase.com) The biggest change in the hosted setup is authentication. Supabase says the hosted MCP server now uses dynamic client registration by default, so users no longer need to manually create a personal access token or an OAuth app before connecting a client. (supabase.com) Instead, the client opens a browser window, the user logs into Supabase, and grants an organization’s access to the MCP client. Supabase says all feature groups except Storage are enabled by default on the hosted server. (supabase.com) The company’s broader OAuth 2.1 server docs show where that flow is heading. Supabase now documents OpenID Connect support, PKCE-based authorization code flows, JWKS, and dynamic client registration for MCP-compatible clients. (supabase.com 1) (supabase.com 2) Self-hosted MCP works differently. Supabase says the MCP server in a self-hosted deployment sits behind the internal API, does not currently offer OAuth 2.1 authentication, and is not intended to be exposed directly to the public internet. (supabase.com) That split leaves Supabase with two tracks in its docs: a hosted path that reduces setup friction for AI tools, and a self-hosted path that keeps tighter network boundaries around local infrastructure. (supabase.com) (supabase.com) For engineers, the practical lesson is simple: Docker now covers the local Supabase stack, and MCP docs now cover how an AI client gets into a Supabase organization without passing around long-lived tokens. (supabase.com) (supabase.com)