Supabase CLI v3.0.0 hit npm on November 2, 2026, and every install channel switched to it within a day: npm, Homebrew, Scoop, the Linux packages, and the supabase/setup-cli GitHub Action. The headline change is one line in Supabase's v3 changelog: "Database commands target your local stack by default."
That's a sensible default for a CLI. It's also a quiet change to every script, CI job and AI agent that ran supabase db push and expected it to land on production.
TL;DR
In v3, supabase db push runs against your local stack unless you pass --remote, even in a linked project. If a CI job starts a local stack to run tests, Supabase's migration guide says a bare db push "succeeds against that stack and never touches your project", so an RLS fix can go green and never ship. Hosted changes now also need consent (--yes in CI). Check your CLI version, grep your workflows and scripts for bare db push, and don't let an agent add --yes for you.
What we saw when we ran it
We installed supabase@3.0.0 from npm, created a project with one migration, and marked it as linked to a hosted project ref. Then we ran the command most deploy scripts and agent transcripts contain:
$ supabase db push
Target: local stack (127.0.0.1:54322) · default; pass --remote for abcdefghijklmnopqrst
... "code":"DbConnectError" ... "suggestion":"Make sure Docker is running, then run: supabase start
To use the linked project instead, pass --remote."
$ echo $?
1
That output is our own observation on v3.0.0, not documented wording. On a laptop with nothing running, it's the good failure. It's loud, it exits 1, and it tells you exactly what changed. The CLI even names the linked project it ignored.
The bad case is when a local stack is running. That's common in CI, because the usual way to test RLS policies is to supabase start, run the tests, then deploy. Under v3 that last db push applies your migration to the throwaway test database, reports success, and the job goes green. Production is untouched.
The migrations that matter most are the ones this hides. A migration that enables RLS on a table, drops an over-broad policy, or revokes a grant fixes something that's exposed right now. If it "deploys" to the CI container instead, your live database stays open and every signal you look at says it was fixed.
Who this affects
You're affected if any of these run the Supabase CLI against your hosted project:
- A GitHub Action using
supabase/setup-cliwithout aversion. Per the migration guide it installslatest, which is now v3, "unless abun.lock,pnpm-lock.yaml, orpackage-lock.jsonat the root of your repository pinssupabase."yarn.lockisn't on that list. As of November 5, the action's own README still shows a release job that runssupabase linkand then a baresupabase db pushon unpinnedsetup-cli@v3, so a workflow copied from it now targets the local stack. - A deploy script or
package.jsonscript withsupabase db push,db pullorgen typesand no--remote. - An AI coding agent (Claude Code, Cursor, Windsurf) that you've let run Supabase commands. Agents learned
supabase db pushfrom years of pre-v3 docs and tutorials.
If you only use the Supabase dashboard, or you build in Lovable or Bolt and never touch the CLI, this doesn't reach you.
The agent problem: --yes
v3 also asks before changing a hosted project, and for anything that changes an existing project the default answer is No. When it can't ask, the command fails with HostedMutationConsentRequiredError and changes nothing. You give consent with --yes or SUPABASE_YES=1. The v3.0.0 help text adds a detail worth knowing: hosted changes only accept that flag or a SUPABASE_YES "exported in your shell, never supabase/.env or another .env file", so an agent can't grant itself consent by writing a dotfile.
The migration guide is explicit about agents: "Agents never get a prompt, even in a terminal," and "the error tells the agent to ask you before it passes --yes."
That's the right design, and it only works if the agent actually asks. The fastest way for an agent to make a failing command pass is to add the flag the error mentions. If you see --yes appear in a command your agent ran, or in a script it edited, that's your production database being changed without you reviewing the migration.
Put the rule where the agent reads it. One line in CLAUDE.md, .cursorrules or your agent's instructions file does it: "Never pass --yes or set SUPABASE_YES on Supabase CLI commands. Stop and ask me instead."
config push can now send your localhost Site URL to production
This one doesn't involve db push at all. Commands that only make sense against a hosted project, like config push, functions deploy and secrets set, still default to the linked project. And in v3, supabase config push sends the SUPABASE_* overrides from .env.local. Supabase's own example: "a SUPABASE_AUTH_SITE_URL in .env.local replaces the site_url in [remotes.production]."
Plenty of projects set that variable to http://localhost:3000 so local magic links and password resets work. Push config from that machine and production auth emails start linking to localhost. Real users click a reset link and land on a page that doesn't exist.
Two smaller auth changes ride along. v2's config push applied [auth.sms] enable_confirmations inverted; v3 applies it as written, so true now really turns phone confirmation on, and phone sign-ups need a working SMS provider. And enable_signup under [auth.email] and [auth.sms] is deprecated in favour of enabled.
The 5-minute check
Find out which CLI you're on
Run supabase --version locally, and check your CI logs for the version setup-cli installed. Anything starting with 3. has the new behaviour.
Grep for bare database commands
grep -rnE "supabase (db (push|pull)|gen types)" \
.github/ package.json scripts/ Makefile 2>/dev/null | grep -v -- "--remote"
Every hit is a command that now targets the local stack.
Fix or pin
Adding --remote alone isn't enough in CI: it turns a silent local push into a consent failure. Supabase's quick reference gives supabase db push --yes --remote, with --remote at the end of the command so it doesn't swallow the next word as a value. If some machines are still on v2, the guide says supabase db push --linked --yes works on both. Only use --yes where the migration was reviewed in the PR. Or pin v2 while you migrate. Supabase lists npm install supabase@v2.stable --save-dev for npm projects and version: v2.stable on supabase/setup-cli.
Check what production actually has
Run supabase migration list --remote and compare it against your supabase/migrations folder. Any migration that's local-only since November 2 didn't ship, whatever CI said.
Check your Site URL
In the Supabase dashboard, under Authentication, then URL Configuration, confirm the Site URL is your real domain. Then move SUPABASE_AUTH_SITE_URL out of any .env.local you push config from, and run supabase config diff before the next config push: it shows every property that's about to change.
Then check the live app, not the migration log
The point of an RLS migration is the effect on your running app: who can read which rows with the public anon key. A migration log tells you what the CLI thinks happened. A request from outside tells you what an attacker sees. That's the check CheckYourVibe runs against your live site, and after a CLI upgrade it's the one that settles whether the fix actually landed.
Why didn't supabase db push update my production database?
If you are on Supabase CLI v3 (released November 2, 2026), db push targets your local stack unless you pass --remote. A linked project is no longer enough. Run supabase --version, and if it says 3.x, use supabase db push --remote, plus --yes in CI where the migration was reviewed.
What is HostedMutationConsentRequiredError?
It's the v3 error a command throws when it would change a hosted Supabase project and can't ask you first, for example in CI or when an AI agent runs it. The command changes nothing. Passing --yes, or setting SUPABASE_YES=1, gives consent.
Should I let my AI coding agent add --yes?
Not by default. Supabase's own guidance is that agents never get the prompt and the error tells them to ask you first. If your agent adds --yes on its own, it's approving changes to your production database without you seeing them.
How do I stay on Supabase CLI v2 for now?
Pin it. Supabase lists npm install supabase@v2.stable --save-dev for npm projects and version: v2.stable on the supabase/setup-cli GitHub Action. Treat it as time to update your scripts, not a permanent setting.
Can config push change my production Site URL to localhost?
Yes, in v3. config push now sends SUPABASE_* overrides from .env.local, so a SUPABASE_AUTH_SITE_URL set to a localhost address for development replaces the site_url in your production remote block. Auth emails then link to localhost.
Did your last RLS fix actually ship?
Scan your live app and see what the anon key can read today, whatever your migration log says.