Supabase Select 2026: Add an MCP Server to Your App in One Command

Supabase launched a BYO-MCP feature at its Select 2026 conference on October 2 that lets any Supabase app expose a Model Context Protocol server to users with a single command, npx shadcn@latest add @supabase/mcp-server, running as an Edge Function with OAuth 2.1 authentication through Supabase Auth and access governed by existing Row Level Security policies. Supabase also shipped OrioleDB into public beta on Pro, Team, and Enterprise plans, reporting roughly 112,800 transactions per minute versus heap's 37,000–65,000 in a TPC-C derived test on an 8XL compute instance (32 vCPU, 128 GiB RAM), a 1.8x improvement, though Supabase does not recommend it for production yet. The company additionally introduced an alpha native local development runtime that runs without Docker on Linux and macOS Apple Silicon, and a declarative schema workflow powered by its pg-delta engine.

Supabase held its Select 2026 conference on October 2, and the headline is not the storage engine benchmarks or the schema tooling — though both are worth your time. The headline is this: every SaaS product built on Supabase can now expose an MCP server to its users, in a single command. AI agents can connect to your product data, authenticate as your users, and operate under your existing Row Level Security policies. That is the shift. The rest is gravy. Your App Gets an MCP Server The new BYO-MCP feature https://supabase.com/docs/guides/ai-tools/byo-mcp lets you add a Model Context Protocol server to your Supabase app so users can connect their own AI agents to your product. The server runs as an Edge Function. Authentication goes through Supabase Auth using OAuth 2.1. Your RLS policies decide exactly what each agent can touch — no per-tool permission code, no custom authorization layer. Add it with one command: npx shadcn@latest add @supabase/mcp-server That scaffolds the Edge Function, wires up the auth flow, and leaves you with an MCP endpoint users can point their agents at. A user opens Claude, Copilot, or Cursor, adds your MCP URL, signs in with their existing account, and their agent can now query your data — scoped to what RLS says they can see. This matters more than it looks. The SaaS products that add MCP support in the next six months will have a feature that feels like a superpower to their power users. The ones that skip it will spend 2027 playing catch-up when their users start expecting every product to be agent-accessible. Add the scaffold to your backlog this week. OrioleDB Is in Public Beta — And VACUUM Is Optional Supabase also shipped OrioleDB into public beta https://supabase.com/changelog/orioledb-public-beta , available on Pro, Team, and Enterprise plans. OrioleDB replaces Postgres heap storage with an undo log, which means dead rows are reclaimed automatically instead of accumulating until VACUUM clears them. Transaction ID wraparound is also gone, replaced by 64-bit transaction IDs. The benchmark numbers are compelling: in a TPC-C derived test on an 8XL compute instance 32 vCPU, 128 GiB RAM, data set larger than memory , OrioleDB hit roughly 112,800 transactions per minute versus heap’s 37,000–65,000. That is a 1.8x improvement at high concurrency, and it holds as connection counts climb. The honest caveat: this is a public beta, there is no SLA, and Supabase does not recommend it for production yet. But if you have a high-write staging environment, or a side project where you want to see what VACUUM-free Postgres actually feels like, now is the time to try it. Create a new project and select OrioleDB at the storage engine step. No Docker? No Problem Alpha Supabase’s local development stack can now run as native processes instead of Docker containers. This means the CLI works on machines without a Docker daemon — CI runners, Claude Code sandboxes, Codex environments, and any other sandboxed context where spinning up Docker is not an option. Enable it in your config: experimental stack = true Or set SUPABASE EXPERIMENTAL STACK=1 in your environment. The native runtime https://supabase.com/docs/guides/local-development/docker-and-native-runtimes runs on Linux and macOS Apple Silicon. You also get multiple independent Supabase instances per directory, so parallel checkouts of the same repository no longer step on each other. This is alpha and off by default — treat it accordingly — but it solves a real problem for teams that hate maintaining a Docker daemon just to run local Supabase. Declarative Schemas 2.0 Supabase’s pg-delta engine now powers a full declarative schema workflow https://supabase.com/docs/guides/local-development/declarative-database-schemas . Define your Postgres schema in SQL files under supabase/schemas/ , run supabase db schema declarative sync , and the CLI generates the migration. Run supabase config pull to sync any dashboard changes back into your files. The repository becomes the single source of truth for both schema and configuration, with no custom DSL or YAML — just SQL. If you have been managing migrations by hand or relying on supabase db diff after every change, this is a cleaner workflow. The schema-as-code pattern that infrastructure teams have expected from Terraform and Pulumi is now available for your database layer. What to Do This Week The priority order is straightforward. First, add the MCP server scaffold to any Supabase project that has paying users or power users likely to use AI agents — it is a one-command addition with immediate value. Second, test OrioleDB on a non-production project if you have write-heavy workloads. Third, enable Docker-free local dev if you work in sandboxed environments, and try the declarative schema workflow if you have been fighting migration drift. Supabase is clearly betting that AI agents become primary clients of SaaS data layers. With BYO-MCP, they have given developers the infrastructure to win that bet early https://supabase.com/blog/supabase-select-2026-recap .