2026 year-in-review essay

The 2026 Supabase changelog is the first BaaS release log written for an agent first.

That is the argument of this page. Read every entry from January through April 2026 and a pattern shows up: every quarter shipped at least one feature that a machine reads more naturally than a person does. Index Advisor, file-based SQL snippets, log drains to Datadog and Sentry, docs over SSH for Claude Code, declarative schema diffs, an MCP server, retries in every client. The features still work for humans. They just stop being human-first.

The other half of the argument is a single file at ~/.fazm/mcp-servers.json, added in Fazm v2.4.0 on April 20. That file is what makes the year of changes legible from one Mac window instead of seven.

Matthew Diakonov, Written with AI

Published April 28, 202611 min

The thesis, stated plainly

Pick any month of the 2026 Supabase release notes and look at what got the screenshot in the announcement post. January was the Stripe Sync Engine, a button you press to materialise a Stripe account inside Postgres. February was a paragraph about saving SQL snippets locally so a team could share them through git. March was Log Drains. April was docs over SSH and an MCP server. Each of those is a feature a person can use. Each of them is a feature an agent prefers.

That is not how every other guide on this writes about it. Most articles march through the entries in date order with a sentence on what each one does, then stop. The reader gets a list. They do not get the through-line. The through-line is that 2026 is the year Supabase started writing release notes that read better to an agent than to a human eye, and that pattern is itself the most useful thing to know if you are building anything that might one day delegate part of its workflow to an LLM with project access.

The rest of this page is the evidence. Then a counterargument. Then what to do with it.

The four quarters as one motion

Watch them in order. Each frame is one Developer Update or one cluster of mid-month entries. The shape repeats: a UI feature people will see plus a hidden one shaped for a machine.

2026 Supabase release notes, in motion

01 / 05

January, v1.26.01

Stripe Sync Engine, PostgREST v14, Python type generation, Index Advisor surfaces. The dashboard becomes a model invocation surface.

The chronology, with the agent-readable bit called out

A vertical read of the year. Each step is one Developer Update plus the mid-month additions. The thing in italics in your head as you read each one: which entry on this list would a human have skipped, and which one would an agent have noticed first.

2026 release notes, ordered

January 2026, v1.26.01: AI behind dashboard buttons

Stripe Sync Engine becomes one-click in Studio, mirroring customers, subscriptions, prices, invoices, charges, and refunds into Postgres so they answer plain SQL. PostgREST v14 lifts GET throughput on the data API. The CLI gains Python type generation, joining the existing TypeScript and Swift generators. Index Advisor surfaces inside the Table Editor with apply-inline suggestions. Read it as: the dashboard is the new place to invoke a model.

February 2026: files, not rows

pg_graphql gets disabled by default on new projects (existing projects unaffected). Edge Functions accept drag-and-drop zip files for whole bundles, including the move from a different project. SQL snippets in Studio start saving locally and a supabase/snippets folder format makes it shareable through git. The pattern is the same in three places: Supabase prefers files an agent or a grep can read over rows in a private database.

March 2026, v1.26.03: streams, not screenshots

March 6 brings a rate limit on recursive Edge Function calls including outbound fetch() between functions. Log Drains land on Pro with first-class targets for Datadog, Grafana Loki, Sentry, Axiom, S3, and custom HTTPS endpoints. Object listing goes up to 14.8x faster on 60M+ row datasets via a hybrid skip-scan algorithm. The Edge Functions dashboard gains inline testing and zip download. Logs and listings stop being a UI screenshot problem and become a stream.

April 2026, v1.26.04: docs piped to Claude Code

Multigres open-sourced as a Kubernetes operator with rolling upgrades, pgBackRest PITR, and OTel tracing. The GitHub integration opens to every plan including free. Supabase joins Stripe Projects as a co-design partner. Docs become reachable over SSH with an explicit Claude Code pipe story. pg-delta ships as a public alpha for declarative schemas (PRs supabase/cli#4966 and #5057). Automatic PostgREST retries on April 20 in supabase-js 2.102.0, supabase-swift 2.43.0, supabase-flutter 2.7.0, supabase-py 2.29.0. RLS Tester preview opens on April 24.

What changes when a release note reads better to a machine

The toggle below is the cleanest illustration of what the year actually changed. Same release notes, two ways of reading them. The human read finishes a list. The machine read finishes a workflow.

A 2026 Supabase release note, two readers

Open the changelog page in a browser. Scroll. Read the headline. Open Studio in a tab to see the new dashboard surface. Read the announcement blog post. Switch to a terminal to bump the client library. Switch back to the editor. Try the new feature once. Close the tab.

List-shaped reading, no compounding context
Each entry visited then dismissed
Cross-window state lost between tabs
No persistent memory of what changed

The receipts: ten 2026 entries that prefer a machine reader

One per checkbox. None of them are agent-only. They all also work for a person. The point is which audience the feature was shaped for.

2026 entries that read more naturally to an agent

Index Advisor inside the Table Editor (January 2026)
AI-described table filters that translate prose into a where-clause
SQL snippets saved to supabase/snippets/ as plain files
Edge Functions zip drag-and-drop, including for migrations between projects
Log Drains to Datadog, Grafana Loki, Sentry, Axiom, S3, custom endpoints
Object listing up to 14.8x faster on 60M+ row datasets
Docs over SSH with an explicit Claude Code pipe in the changelog
pg-delta declarative schema diffs (supabase/cli#4966, #5057)
supabase-community/supabase-mcp server for MCP-aware clients
Automatic PostgREST retries in all four official clients on April 20

Anchor: the file Fazm v2.4.0 added on April 20

This is the file. Drop it at ~/.fazm/mcp-servers.json or open Fazm Settings, MCP tab, and the Supabase MCP server is registered. The shape is the same shape Claude Desktop uses, which is deliberate. A config that works in one MCP-aware client should work in another without rewriting.

// ~/.fazm/mcp-servers.json
// File added in Fazm v2.4.0, 2026-04-20
{
  "mcpServers": {
    "supabase": {
      "command": "npx",
      "args": [
        "-y",
        "@supabase/mcp-server-supabase@latest",
        "--access-token",
        "<YOUR_SUPABASE_PAT>"
      ],
      "env": {}
    }
  }
}

The release entry lives in CHANGELOG.json in the Fazm repository under v2.4.0, dated 2026-04-20: "Added custom MCP server support via ~/.fazm/mcp-servers.json with Settings UI to add, edit, and toggle servers." The Supabase MCP server itself lives at github.com/supabase-community/supabase-mcp. Both are open source.

Why the date is not coincidence-shaped

April 20, 2026 is the calendar square that gives the year its punchline. On that single Monday, Supabase shipped automatic PostgREST retries in supabase-js 2.102.0, supabase-swift 2.43.0, supabase-flutter 2.7.0, and supabase-py 2.29.0. The same Monday, Fazm v2.4.0 shipped ~/.fazm/mcp-servers.json and the Settings UI to manage MCP servers. The platform side of the database now retries on its own. The desktop side now has the file format that lets a Mac assistant talk to the database without a browser tab.

The two changes are independent. They are also the same change, shape-of-things-to-come edition. A workflow that used to require a person clicking around in Studio with a credentials manager open in another window can sit inside one assistant conversation, with retries, with structured project access through MCP, with the docs cat'd in over SSH, and with the accessibility tree of every other window open on the laptop available as additional context.

2026 surfaces a Mac assistant reads at once

Studio in ChromeEdge Functions dashboardsupabase-js 2.102.0 retriessupabase-swift 2.43.0supabase-flutter 2.7.0supabase-py 2.29.0Datadog log drainSentry log drainsupabase/snippets folder in VS Codepg-delta plan in TerminalClaude Code piping ssh docs~/.fazm/mcp-servers.json

The counterargument

The fair pushback is that Supabase did not stop building for humans in 2026. The Schema Visualiser got clickable relations and context menus in April. Fix with Assistant lives inside a browser tab and is meant to be looked at by a person. The new Edge Functions dashboard with inline testing in March is a UI win, not an agent win. PostgREST v14 helps every existing app whether or not anyone is reading the changelog. The GitHub integration opening to free in April matters most for the solo developer with one Postgres and a hobby project, who is exactly the human Supabase was always built for.

That is true and it does not undo the argument. The claim was never that 2026 dropped human users. It is that 2026 stopped treating the human path as the only path. Of the entries shipped between January and April, roughly half make sense whether or not an agent is in the loop, and roughly half are noticeably more useful when one is. The mix shifted. That is the year-over-year delta that previous years' release notes did not have.

The clean test: pick a 2025 Supabase release note and ask whether a Mac assistant with project access would treat it differently than a person reading a blog post. Most of the time the answer is no. Then pick a 2026 release note. Index Advisor, supabase/snippets, log drains, docs over SSH, pg-delta, the MCP server. The answer flips on those.

The shortest version

2026 was the year Supabase started writing release notes that read more cleanly to a machine than a human. Every quarter shipped at least one feature that prefers a machine reader. April 20 is the punchline: retries in every official client and a Mac assistant config file with the same MCP shape Claude Desktop uses, on the same calendar day.

If you are pinning client libraries this week, pin to the April 20 cut. If you are running a Mac, drop the supabase MCP entry into ~/.fazm/mcp-servers.json and let the assistant read across windows. If you are reading future Supabase release notes, watch for the entry that does not look exciting in a browser. That is probably the one that just got shipped to be parsed.

See the 2026 Supabase + MCP loop on a real Mac

Fifteen minutes, screen-shared. Bring a Supabase project. Watch a desktop assistant pipe the docs over SSH, query the live schema through the MCP server, and read your editor and Datadog tab in the same conversation.

Questions about the 2026 Supabase release notes

What did Supabase ship across 2026 so far, in one paragraph?

January (Developer Update v1.26.01): the Stripe Sync Engine became a one-click install in Studio, the data API moved up to PostgREST v14, the CLI gained Python type generation, and Index Advisor surfaced inside the Table Editor. February: pg_graphql was disabled by default on new projects, Edge Functions added drag-and-drop zip uploads for whole bundles, and SQL snippets in Studio started saving as files locally so a team could share them through git in a supabase/snippets folder. March (Developer Update v1.26.03): a new rate limit on recursive Edge Function calls landed on March 6, Log Drains opened on Pro to Datadog, Grafana Loki, Sentry, Axiom, S3, and custom endpoints, object listing got up to 14.8x faster on datasets above 60M rows via a hybrid skip-scan algorithm, and the Edge Functions dashboard added inline testing and zip download. April (Developer Update v1.26.04): the Multigres Kubernetes operator was open-sourced, the GitHub integration opened to every plan including free, Supabase joined Stripe Projects as a co-design partner, docs became reachable over SSH with an explicit Claude Code pipe, pg-delta declarative schemas shipped as a public alpha, automatic PostgREST retries landed in supabase-js 2.102.0, supabase-swift 2.43.0, supabase-flutter 2.7.0, and supabase-py 2.29.0 on April 20, and the RLS Tester preview opened on April 24.

What is the through-line of the 2026 changelog if I read it cold?

Every quarter shipped at least one feature that prefers a machine reader over a human operator. January: Index Advisor and AI table filters put model-driven suggestions behind dashboard buttons. February: SQL snippets stopped being database rows and started being files in supabase/snippets/, which is what an agent or a grep can read. March: Log Drains formalised log streams into shapes Datadog and Sentry already parse. April: docs over SSH explicitly endorses piping into Claude Code, pg-delta makes the schema declarative the way an agent expects, and the Supabase MCP server opens a direct query channel to any AI client. None of these are agent-only features. They all also work for humans. The point is that the human path is now the secondary one in each release.

Which version pins do I actually have to know to take the 2026 changes seriously?

supabase-js v2.102.0, supabase-swift v2.43.0, supabase-flutter v2.7.0, and supabase-py v2.29.0 are the April 20 cut that opted everyone into automatic PostgREST retries on transient errors (up to three attempts on GET and HEAD, exponential backoff, mutations never auto-retried). PostgREST v14 is the platform-side bump from January, with roughly 20 percent more requests per second on GETs, no client code change needed. supabase/cli pull requests #4966 and #5057 are where pg-delta lives. supabase-community/supabase-mcp is the MCP server. Fazm v2.4.0 (2026-04-20) is the desktop counterpart with custom MCP server support. The CHANGELOG.json file in the Fazm repo is the canonical record for that.

Why does this page bring a Mac app into a database changelog conversation?

Because the 2026 Supabase releases stopped fitting in one window. Studio is a browser tab. The Edge Functions dashboard is another browser tab. Claude Code is a terminal. pg-delta is a CLI on disk. Log Drains land in Datadog or Sentry, which are their own browser tabs. The Supabase MCP server is a process. Each release does its job inside one of those surfaces, but the work a person actually does crosses three of them per task. A consumer Mac assistant that reads the macOS accessibility tree of whatever app is under the cursor is the cheapest way to hold all those surfaces in one frame. Fazm is that assistant, and v2.4.0 added a config file that registers MCP servers, including the Supabase one, on the same calendar day Supabase shipped retries in every client.

What is Fazm v2.4.0 and why does the date matter?

Fazm v2.4.0 shipped on April 20, 2026. The headline change is custom MCP server support via a config file at ~/.fazm/mcp-servers.json, with a Settings UI to add, edit, and toggle servers. The shape of the file is the same shape Claude Desktop uses, which means the supabase-community/supabase-mcp server drops in without rewriting. The date matters because that exact Monday is also when Supabase released supabase-js 2.102.0, supabase-swift 2.43.0, supabase-flutter 2.7.0, and supabase-py 2.29.0 with retries on. Two products, same calendar square, both unlocking the same workflow: a Mac assistant that can both query the project through MCP and watch any window on the laptop without taking screenshots.

Did Supabase ship anything in 2026 that breaks existing code?

Two things to watch. First: pg_graphql is no longer enabled by default on new projects starting February 2026. Existing projects keep it. New projects that need GraphQL have to opt in. Second: the March 6 rate limit on recursive Edge Function calls is a behavior change for code that fans out from one function to many others, including outbound fetch() calls between functions. If your architecture relies on that, the limit can become visible. Everything else in the 2026 list is additive, opt-in, or a UI change.

What does pg-delta change in practice for someone already using migrations?

pg-delta lets you keep declarative SQL files for your schema and have the CLI generate migrations from the diff between your declared end state and the live database. Postgres 15 and newer only, public alpha, interactive and non-interactive modes, and the engine warns on destructive changes before applying them. The two pull requests to read are supabase/cli#4966 and #5057. If you already have a stable migration flow it is not a forced upgrade. If you are migrating new schemas often or generating them programmatically, the declarative path is closer to how an agent would prefer to write schema changes anyway.

How does Log Drains in March 2026 change observability on Supabase?

It opens five real targets and a custom HTTP endpoint to anyone on the Pro plan: Datadog, Grafana Loki, Sentry, Axiom, S3, and a generic webhook. Postgres logs, Auth logs, Storage logs, Edge Functions logs, and Realtime logs can all be drained. Before March 2026 the only way to ship logs out was the Logs UI plus manual export. After March, an alerting setup that already exists for the rest of your stack picks up Supabase events without bespoke code. For a desktop assistant this matters because the same Datadog dashboard a human reads is now reachable to an MCP-aware agent that can hop from a Studio tab to a Datadog tab via the accessibility tree.

Where is the authoritative source for any single 2026 Supabase release note?

supabase.com/changelog is the live surface. The Developer Update releases are tagged in github.com/supabase/supabase: v1.26.01, v1.26.03, v1.26.04 are the Jan, March, and April tags respectively. Each official client library has its own GitHub release page (supabase-js, supabase-swift, supabase-flutter, supabase-py) with permanent version tags. supabase-community/supabase-mcp is the MCP server repo. supabase/stripe-sync-engine is the Stripe sync repo. supabase/cli holds pg-delta. If anything on this page drifts from those sources, those sources win.

If I am not on a Mac, does any of this matter?

Yes for the Supabase half, no for the desktop-MCP half. The version pins, retry behavior, PostgREST upgrade, Log Drains, pg-delta, the GitHub-on-free move, and the Supabase MCP server are all platform-agnostic. Fazm is macOS-only because it depends on the AXUIElement accessibility API. The MCP angle is not Mac-specific in principle: any MCP-aware client (Claude Desktop, IDE plugins, custom CLIs) can connect to the Supabase MCP server. This page is written from the Mac side because Fazm v2.4.0's MCP config landed on the same Monday as the four Supabase client cuts, which makes the connection observable rather than hypothetical.

Adjacent 2026 release coverage