Convex vs Supabase for AI agents

As of 16 Oct 2026, Convex gets an AI agent further: L3 · Does the work against L1 · Needs a person.

ConvexL3Does the work
  • Find: Docs for agents: llms.txt, an OpenAPI and an MCP server.
  • Access: Runs locally: Can't sign up for one.
  • Use: 70% of the work.
  • Pay: Payment not reached.
  • Leave: Leaves on its own: Nothing to cancel, made nothing to delete.

Got in; payment not reached. Next: let an agent sign up for an account on its own (without one it does 70% of the core work)

Gets in on its own, L3 · Does the work: a documented no-account route provisions a local deployment, where our agent pushed schema and functions, inserted and patched documents, ran an HTTP action, a cron and a scheduler, and set env vars, each checked back. The cloud side stopped at sign-up: a visible Cloudflare "Verify you are human" challenge blocked account creation, so no cloud project, deploy key or MCP session followed. Core backend work, plus file storage, text and vector search, custom-JWT auth and live subscriptions, ran on the local deployment; cloud paying and account deletion stayed out of reach, since both sit behind that same human checkpoint and the dashboard.

Gets furtherDriven 2 Oct 2026 · The drive and its evidence
SupabaseL1Needs a person
  • Find: Docs for agents: llms.txt, an OpenAPI and an MCP server; a docs page for each core task.
  • Access: Signs up sometimes.
  • Use: 100% of the work.
  • Pay: Needs a person to buy: A CAPTCHA at checkout.
  • Leave: Nothing to cancel: Deletes its data, but the login stays.

Got in; first wall at pay (blocked). Next: let an agent get in on its own in every drive

Supabase is a backend platform for databases, user accounts, file storage and serverless functions. Once the agent signed in with GitHub, it could do almost everything on its own through code: create a database, add users, store files, run functions, and more. It stopped at the very first step: creating an account by email showed a puzzle test it couldn't solve, so it had to use a GitHub sign-in instead, and upgrading to a paid plan hit the same kind of puzzle.

Driven 16 Oct 2026 · The drive and its evidence

The two summaries: written by our agents from each product's drive. How we write.

Where each one stops.

ConvexGot in; payment not reached

SupabaseGot in; first wall at pay (blocked)

Convex and Supabase, checkpoint by checkpoint
CheckpointConvexSupabase
DocsHeadlessHeadless
Sign upNot neededIn a browser
VerifyNot neededNot needed
KeyNot neededIn a browser
First callHeadlessHeadless
First actionHeadlessHeadless
PayNot reachedBlocked
CancelNot neededNot needed
DeleteBlockedBlocked

Headless or in a browser.

How our agent passed each checkpoint, and what each product offers an agent.

Checkpoints by route, API, MCP, CLI and SDK
Checkpoints by how our agent passed them, and what each product offers an agent
ConvexSupabase
Headless (checkpoints)33
In a browser (checkpoints)02
Needs a human (checkpoints)00
Blocked (checkpoints)12
Not tested (checkpoints)10
Work done70% of the core100% of the core
APIYesYes
OpenAPIYesYes
MCP serverYes (official)Yes (official)
CLIYesYes
SDKYes–
llms.txtYesYes

Bot checks, and where a person was needed.

Walls our agent met
ConvexSupabase
Bot checkCloudflare human check on the sign-up page; it got in another wayA challenge on the form stopped one way in (Email and password)
Wall typeA check of our network and browserA challenge on the form
First wallGot in; payment not reachedGot in; first wall at pay (blocked)
A person helpedNoNo

A check of our network and browser: It judged where our agent connects from and its browser, not a person. An agent that works on its own from a data center doesn't get through unless the owner lets signed agents in. How agents in a person's own browser fare isn't measured yet. A challenge on the form: It asks every visitor to prove they are human. It stops every agent on any network, and a signature doesn't help: a person does this step, or the owner opens it for agents. Where we drive from.

Level and part scores.

Convex and Supabase: level and part scores, 0 to 100
ConvexSupabase
LevelL3 · Does the workL1 · Needs a person
Holds upNot measured82 of 100, fair
Agent-ready setup100 of 100, good100 of 100, good
StandardsNone recordedMCP spec of 28 Jul 2026 (current) · MCP discovery (current) · MCP client registration (behind) · OpenAPI 3.0.0 (2 versions behind)
SecurityNo review in this driveReported privately to the owner

Docs match is measured in an Audit. What the scores mean.

Pay, cancel and delete.

Paying, cancelling and deleting the account
ConvexSupabase
PayPayment not reachedNeeds a person to buy
Free plan or trialFree planFree plan
CancelNot neededNot needed
Delete the accountBlockedBlocked
How each one takes payment
  • Convex: Payment not reached.
  • Supabase: A person has to buy: a CAPTCHA at checkout.

Convex vs Supabase: the short answers.

Which can an AI agent use, Convex or Supabase?

Convex gets an AI agent further: L3 · Does the work against L1 · Needs a person. With Convex our agent got in (payment not reached); with Supabase it got in and met its first wall at pay (blocked). Driven 2 Oct 2026 (Convex) and 16 Oct 2026 (Supabase).

Can an AI agent sign up for Convex and Supabase on its own?

Convex: not needed. Supabase: yes, in a browser, but a later step stopped it.

Do Convex and Supabase have an MCP server?

Convex: yes (official). Supabase: yes (official).

Can an AI agent delete its Convex or Supabase account?

Convex: blocked. Supabase: blocked.

The drives and the category.

Put your own shortlist side by side.

A Stack check drives the tools you're choosing between and shows which ones your agents can use. $490 for three tools, ready in 24 hours.