Neon is built to be operated by AI coding tools. One command sets up the current directory with agent tooling, a linked Neon project, and a neon.ts config, and from there your assistant can create branches, run migrations, deploy functions, and enable Auth or the Data API on its own (neon init).
What your assistant gets
- An MCP server. The Neon MCP Server lets an assistant manage projects, run queries, and make schema changes through Model Context Protocol tools. Set it up with
npx neon@latest mcp. - Editor plugins. Neon publishes plugins for Claude Code, Cursor, Codex, and others that bundle the MCP server and the core Postgres skills (agent skills). For Functions, Object Storage, and AI Gateway context, run
npx skills add neondatabase/agent-skills -y. - A CLI designed for agents. Every Neon CLI command supports
--output json, and settingNEON_API_KEYauthenticates non-interactively. Agents bind a directory to a project withneon link's non-interactive flags. - Docs as markdown. Any docs page returns markdown when you append
.mdto the URL, andhttps://neon.com/docs/llms.txtindexes every page.
The branch-first loop
The workflow that keeps an agent from corrupting shared state is one Neon branch per feature. neon link binds the directory once; neon checkout <branch-name> creates or switches to a branch and pulls its DATABASE_URL into your .env. A branch is a copy-on-write clone of the parent, so it starts with real data and no copy step (branching). Set a TTL on dev-* branches in neon.ts and they delete themselves (neon.ts).
neon init # agent tooling + linked project + neon.ts
neon checkout dev-add-search # isolated branch, env pulled automatically
neon deploy # provision declared services and functionsWhen there's no account yet
If an agent is building for a user who isn't around, Claimable Neon provisions a project now and hands over a claim link. Unclaimed projects expire in 72 hours and are capped at 100 MB of storage and 1 GB of transfer. Agents fetch https://neon.com/auth.md to discover the protocol.
Scale to zero suits agent-generated apps
Most apps built in an afternoon sit idle afterward. Compute suspends after 5 minutes without queries, so an experiment that never ships costs storage only (scale to zero). The Free plan includes 100 projects, so each experiment can have its own.
How other options compare
- Supabase: publishes an MCP server (public alpha per the feature status table) and an agent skill (
npx skills add supabase/agent-skills). Its own guidance is to point the MCP server at a development project, not production, and to use read-only mode for real data (MCP), because there's no cheap copy to experiment on: branching isn't included on Free, and on paid plans each branch is an hourly-billed environment rebuilt from migrations without production data (branching usage). The Free plan allows 2 active projects and pauses them after a week of inactivity (pricing), which caps an agent that creates a project per experiment. Free projects have no backups (backups), so a wrongDROP TABLEhas no platform-level undo; Neon Free restores to any point in the last 6 hours. And because a Supabase app reaches the database from the client, every exposed table needs a correct RLS policy before the app is safe to share (going into prod, Free plan comparison). - Firebase: Firestore is a NoSQL document database (Firestore); agents generating SQL schemas and migrations target Postgres more directly.
Vendor details verified on 2026-09-02 against the linked pages.

Run neon init and let your assistant take it from there.








