Neon

A deep dive into how we scale Postgres in real time

Three new Neon CLI commands — neon mcp, neon skills, and neon plugins — install Neon into coding agents like Cursor and Claude Code without manual config. The MCP server's inspect_database tool now reports queries stalled over 30 seconds, and new NAT gateway IPs in us-east-2 require allowlist updates.

A deep dive by Neon argues for treating Postgres' write-ahead log as the source of truth, backed by S3 object storage, with the LSN making the timeline addressable for agent-style history operations like branching, point-in-time restore, and parallel isolation.

Lakebase Search extensions can now be installed directly with two CREATE EXTENSION statements, and both extensions were updated to newer versions. The Neon CLI stores credentials in the OS keyring with --keyring and inspects every database on a branch at once; the Neon MCP server is now an official Claude connector.

New guide explains how open-weight models deliver performance on Neon AI Gateway, which calls Databricks Foundation Model APIs for Databricks-hosted open-weight models.

Neon AI Gateway lets you call frontier and open-weight models from within Neon, replacing per-lab API keys with one Neon credential and unified billing at provider list prices. Models are hosted by Databricks on its Foundation Model APIs, and each Neon branch gets its own gateway host with credentials that follow branch lineage.

Neon Functions brings Node.js 24 compute to your branch, colocated with Lakebase Postgres in the same region, with DATABASE_URL and other Neon backend credentials injected automatically. Long-running across requests, it supports WebSockets and SSE via Postgres LISTEN/NOTIFY fan-out, and functions branch with your database with their own public HTTPS URLs per branch.

vercel integration add neon now provisions a Neon database on a Vercel project, injects connection variables like DATABASE_URL, and automatically installs Neondatabase/agent-skills from skills.sh so agents know how to use the provisioned setup. Preview deployments get isolated Neon branches; cleanup follows the integration path's retention policy.

Electric and its PGlite sync engine join Neon, bringing real-time, local-first Postgres sync to the platform. The Neon AI Gateway is now a community provider for the Vercel AI SDK with a single credential for 42+ models, and Neon's agent skills/MCP server ships as an Agent Plugins 1.0.0 package for any conforming client.

The Console sidebar now groups all branch-level backend services at the same level, with project-level items separated from branch-level items; branch Overview lists service status. AI Gateway adds Kimi K3, GLM-5.2, Inkling, and new Gemini/GPT models. Project-level permissions now support existing orgs, and the CLI adds api-keys and profile commands for scoped credentials and multi-account switching.

Neon posits that RL post-trained open-source models can match frontier-model search accuracy for a fraction of the cost, pointing at Castform's post-training pipeline running against Lakebase Search on Neon as the proof.

Neon now supports per-project permissions with four org-level roles (Admin, Editor, Viewer, Collaborator) and project-level Admin, Editor, and Viewer grants. The existing project collaboration functionality is deprecated and replaced by the new Collaborator role plus project-level permissions.

Project-level permissions let you grant Viewer, Editor, or Admin access on individual projects, while the new Collaborator organization role provides no default access until specific projects are granted. Existing organizations are migrated automatically with no action required. Also includes new guides for building durable workflows and an LLM proxy on the Neon backend, Lakebase Search case study and guide, renamed spending limits to spending notifications, schema-qualified table names in the MCP server, and a Drizzle Studio 1.5.1 update with multi-column sorting and keyboard shortcuts.

The Neon CLI now supports neon inspect db for running read-only Postgres diagnostics, and backend log querying is available.

Blog post recounts using the new backend suite (beta) to build a working Git host during a GitHub outage, with no prior knowledge of Git server internals.

New neon inspect db command provides 14 read-only, curated diagnostic queries against Postgres catalog views, covering table/index sizes, bloat, unused indexes, sequential scans, long-running queries, locks, query statistics, vacuum status, replication slots, and local file cache hit rate. Connections are resolved automatically for linked Neon projects, and results can be output as table, JSON, or YAML.

Neon Object Storage, Functions, and AI Gateway are now in beta, and a new TypeScript SDK (@neon/sdk) replaces @neondatabase/api-client as the recommended API client. Passkey sign-in is available, and the CLI adds a git-style diff command alongside config diff output.