Supabase Compute runs AI agents inside a project, next to Postgres

Supabase has opened a private alpha of a product called Supabase Compute. It runs long-lived services next to a project's Postgres database, in the same project and the same region. The announcement is dated 27 October 2026, by engineers Lakshan Perera, Nyannyacha, Matt Johnston and designer Francesco Sansalvadore.

Supabase says the gap it fills is work that does not fit in an Edge Function. An Edge Function has a 400-second limit. It also has a small memory budget. And it runs JavaScript only. An AI agent that calls a model many times has to live on a second platform. So does a durable job that processes a queue for hours. So does a service written in another language. That second platform has its own deploys, its own logs and its own bill.

What runs on it

A Compute service gets a full Linux environment. Supabase lists five use cases on the launch page. An AI agent that spins up and coordinates many tasks. An API or web service in any language. A durable job processor that pulls tasks from Postgres. A data sync or embedding backfill that runs for hours. A WebSocket service for voice agents and real-time apps. Services deploy as Node, Deno or any Dockerfile. The slot sizes are 2 GB of memory with 1 vCPU, or 4 GB with 2 vCPUs.

Each service picks public or private. Public services get a URL and suspend when idle. Private services have no URL and keep running. Every service inherits the project's secrets, and its logs appear in the dashboard. The CLI is supabase compute, from Supabase CLI v2.120.0 or later. There is also an alpha Management API to deploy, list and delete services.

The latency figure Supabase quotes

Supabase names one customer, Dreambase, which moved a data agent onto Compute to test it. Supabase says Dreambase measured median Postgres query times of 2.9 ms from the agent on Compute. The same queries ran at 65 ms from another region. The figure is Supabase's own, from one customer's test. No measurement method is given. The number is what Supabase says, not a general promise.

What the alpha does not do

Supabase states the limits on the launch page. The alpha runs in us-west-2 only. A service can be restarted at any time, so long-running work has to save state to Postgres. There is no autoscaling, no GPU and no persistent disk. Public services suspend when idle; private services keep running. Supabase also says that sandboxes for AI agents, where the agent runs untrusted code, are not part of the alpha. People who want them have to ask on the waitlist.

Supabase is open about calling this an evaluation release, not a production one, and about the shape changing. A team that wants to try it joins the waitlist with a Project ID. The announcement sits inside a wider Supabase release. The same batch also included a declarative schema workflow and an alpha of a local backend that runs without Docker.

Source

Supabase Compute: run your agents inside your project, on the Supabase blog.