Neon Neon's new TypeScript SDK is built as much for agents as for people
Neon shipped @neon/sdk, a zero-dependency TypeScript client for its control-plane API that folds multi-step provisioning into single calls. The framing points at platform builders and automated callers, not just scripts written by hand.
Neon has released @neon/sdk, a rewritten TypeScript client for the API that provisions and manages its serverless Postgres. It replaces the company’s older auto-generated client with a zero-dependency, fetch-based library, and it wraps the fiddly multi-step parts of driving a control plane, spin up a project, wait for it to be ready, hand back a connection string, into single calls. The framing is the tell: Neon is pitching this at automated callers and platform builders as much as at people writing scripts by hand.
What shipped
The new SDK succeeds @neondatabase/api-client, which Neon describes as auto-generated and published straight from its private cloud repo. The replacement is generated from the same OpenAPI spec but adds an ergonomic layer on top, ships with zero dependencies, drops the older axios requirement, and is ESM-based with better error handling.
That ergonomic layer is where the work is. A createAndConnect call provisions a project, polls until the underlying operations finish, and returns a ready-to-use connection string in one step. There is snapshot restoration with an optional preview-before-commit, project transfers across organizations, and branch creation that provisions compute automatically. Automatic retries and readiness polling are handled inside the library, and the raw API endpoints are still available as tree-shakeable functions for anyone who wants to drop below the abstractions. The older client is not being deprecated any time soon, so this is an additive move rather than a forced migration.
For a developer driving Neon by API, the win is specific: the annoying parts, polling for is-it-ready-yet, stitching multi-step flows together, and handling transient failures, move out of your glue code and into the SDK. That is the difference between a control plane you can script against comfortably and one you wrap in a pile of custom retry logic.
Where this lands
Neon aims this squarely at developers and agents building platforms on top of it: CI/CD pipelines, automation scripts, and the products that already provision Neon under the hood, including Replit, Netlify DB, and Vercel’s marketplace integration. That audience is not incidental. Neon’s core differentiator has always been that databases are cheap to create and branch, and that only pays off if creating and branching is fast and programmable. An SDK with readiness polling and a one-call provision-and-connect is the tooling that makes a database per preview, per tenant, or per agent run practical rather than a custom polling loop each team rebuilds.
It also fits where Neon is headed. Since being acquired by Databricks, the company has leaned into being the Postgres substrate other products provision on top of, and an ergonomic control-plane SDK is a small but on-strategy piece of that: the easier it is to call, the more platforms will call it. The agent angle rhymes with a broader pattern this month, where automated clients increasingly provision their own infrastructure. A database an agent can create, connect to, and later tear down in a single typed call sits comfortably alongside deploy targets moving the same way.
What’s worth watching
The near-term friction is compatibility. A zero-dependency, ESM-only client is the modern choice, and it will still be a speed bump for teams anchored in CommonJS, so how cleanly it drops into older toolchains is worth a look before you commit. The more interesting question is how far Neon pushes the agent-provisioning story into first-party primitives, ephemeral branches scoped to an agent run with automatic teardown would be the natural next step, and whether the docs and examples steer new projects to this client over the one it replaces. Adoption relative to the old API client is the signal to track. Stackmaven’s follow-up coverage will land on or around October 12.