Supabase Select puts coding agents under Auth and RLS while scale stays gated

Coding agents already write the application. The backend is still the half that lives in dashboards: schemas clicked into existence, settings that never land in git, work that outlives an Edge Function, and no clean way for a user’s ChatGPT or Claude session to act inside the product under the policies you already wrote.

On 2 October 2026, at Select 26 on San Francisco’s 20th Street, Supabase spent a developer day on that gap. Paul Copplestone and Ant Wilson keynoted. Matching product posts went up the same morning. The funding round and Turso deal from that day sit on Frontier’s earlier Dispatch. This Stack is what landed in developers’ hands, and what stayed behind a waitlist.

Declarative Schemas, config-in-code, and app MCP put the backend where agents already work

Declarative Schemas 2.0 makes SQL schema files the source of truth. pg-delta, the CLI’s new diff engine, writes the migration from those files. Auth providers, API limits, storage buckets, and Compute definitions live in config.toml.

supabase config pull and supabase pull bring dashboard drift back into the repo. New projects from supabase init default to pg-delta. Existing ones opt in through config.

Native local mode runs the stack as ordinary processes. It starts inside Claude Code sandboxes, Codex, Perplexity Computer, and CI runners with no Docker daemon. Each checkout can keep its own stack. The mode is alpha and off by default. Stage claims of nine-times-faster starts stay attributed talk until the docs publish measured figures.

Supabase Compute runs long-lived Linux services next to Postgres: any language, Dockerfile deploys, configurable CPU and memory, no Edge Function timeout. Product materials describe agent sandboxes and always-on HTTP workers beside the database. Services are defined in the same config file. Compute is private alpha. Edge Functions keep running unchanged. Treat idle billing and ten-millisecond access claims as waitlist-doc language.

The sharpest surface for agents-as-users is the app MCP server. It is separate from the official Supabase MCP used to build a project. One library command adds an authenticated Edge Function so a user’s agent signs in through your Auth flow and only sees rows your RLS policies allow.

Supabase says it already uses the pattern so ChatGPT can sign into the dashboard. Treat that as a vendor demo. The security model is the point: the LLM becomes another principal under policies you already wrote.

Operate tooling lets agents watch production and still leaves a human on the merge

query_logs gives agents SQL access to project logs. Health Check Advisors flag elevated error rates on the Data API, Auth, Storage, and Edge Functions. Database Connections shows locks and long sessions and lets you cancel work.

Explorer, rolling out through mid-October as a Feature Preview, replaces the old SQL editor. It can query logs and the database together and save the run as a Notebook. Notebooks push and pull with the CLI.

Supabase published scheduled prompts for health, security, performance, and resource checks, meant for Claude, Codex, or Cursor against a read-only MCP connection. The loop they describe is observe, test a fix in an isolated environment, then open a pull request. You wake up to the PR. You still decide what merges.

Permissions shipped the same day. Okta-managed auth for the Supabase MCP is generally available, with short-lived tokens bound to the signed-in member. Scoped personal access tokens are the dashboard default for new tokens; start read-only.

MCP elicitations ask for confirmation before creating a paid project or branch, or before SQL that could delete data. The docs say the prompt helps and that the client must support elicitations for it to appear. Without client support, the confirmation never shows.

Pipelines, in public alpha on Pro and above, stream Postgres to BigQuery, ClickHouse, DuckLake, or Snowflake so analytics leave the primary.

OrioleDB is public beta; Multigres and Compute remain invite-only

The scale story at Select was three different ceilings.

OrioleDB is a public-beta storage engine that replaces Postgres heap with an undo-log design aimed at bloat, VACUUM overhead, and transaction-ID wraparound. You pick it at project creation and can mix OrioleDB and heap tables.

Supabase’s TPC-C-derived benches on dbarena.com claim up to 1.8 times heap throughput. Keynote language that rounded that to twice as fast stays on the stage. Same SQL and extensions. The engine’s creators, including Alexander Korotkov, work at Supabase.

Multigres is the multi-node layer from people behind Vitess, including Deepthi Sigireddi and Vitess co-creator Sugu Sougoumarane on the Select bill. It bundles pooling, automated failover, and a path toward sharding. Clusters run three nodes across availability zones.

Multigres is open source as a project and private alpha on Supabase by invite. The company says it is not ready for production workloads. An internal claim of one thousand clean sequential failovers is vendor testimony.

Turso is the horizontal chapter from the same morning: millions of file-backed SQLite databases on a multitenant server, loaded on demand and suspended when idle. Glauber Costa joins to lead agentic infrastructure. Pekka Enberg and the Turso team come with him. Presentation talk of a lighter Spark instance is naming from the stage. Deal terms and the raise stay on the capital Dispatch.

dbarena.com is the trust layer for the performance talk: open methodology, downloadable raw data, and cost-normalised comparisons that launched with Supabase, Amazon RDS, and Google Cloud SQL.

Schema-in-git and app MCP are usable now; Multigres is still research access

Teams whose agents already edit application code can pull schema and config into the repo this week. Teams whose users will ask ChatGPT or Claude to act inside a product can add the app MCP server where Auth and RLS already exist.

Operators who want agents on the watch can enable Health Advisors and Connections, issue a read-only MCP token, and run the published prompts against a staging project. Explorer is worth flipping when the Feature Preview reaches you. Compute and Multigres belong on the waitlist as research access until the alpha labels move. OrioleDB is fair to try on a new project if you can live with beta. Turso matters when you need millions of small databases.

Talk metrics and alpha gates still bound what Select can promise

Select’s marketing line is that the backend is finally agent-native. The maturity table is colder.

Compute and Multigres are private alpha. Native local mode is alpha and off by default. OrioleDB and Pipelines are beta or public alpha. Elicitations reduce accidental spend and destructive SQL; they leave the agent unsafe if the client ignores them or the token is over-scoped.

Mixing up the official Supabase MCP with your app’s MCP server grants the wrong blast radius. Idle Compute billing still has to be read from the waitlist docs. Stage numbers (nine-times-faster native starts, two-times OrioleDB, one thousand clean failovers) stay attributed talk until someone else reproduces them.

What holds without the stage numbers is narrower. Supabase spent a developer day making schema, config, runtime, and permissions addressable to the same tools that already touch application code.

Read the maturity table before you join another waitlist

The lasting Select move is agent-addressable backends under Auth and RLS, with most of the hard scale still gated. Pull schema into git and add an app MCP where you already have Auth. Treat invite-only Compute and Multigres as research until the alpha labels come off.