August 2026 Summaries

Lakebase Postgres, based on Neon’s architecture, autosizes database compute in real time by separating stateless query execution from durable storage, allowing virtual machines to resize, move, start, or stop without relocating database data or dropping connections. Its autoscaling algorithm selects the largest capacity target indicated by CPU load, memory pressure, and the compute cache working set, within user-configured limits; CPU and overall memory are sampled every five seconds, Postgres memory is monitored every 100 milliseconds, and cache demand is evaluated every 20 seconds. To capture workloads that suffer from cache misses despite low CPU use, the system uses a time-aware modification of HyperLogLog to estimate recently accessed database pages across windows from one to 60 minutes, identify when prior heavy activity has ended, and project near-term cache growth. Resizing is coordinated by node-level autoscaler agents, an in-VM monitor, a modified Kubernetes scheduler, and NeonVM infrastructure using QEMU and KVM, which can adjust CPU and memory on live virtual machines or migrate them when needed. The approach enables both rapid scaling up and safe scaling down, with safeguards ensuring that memory is not removed when Postgres still requires it, and production databases reportedly change size more than 32,000 times per month on average.

Neon has launched Neon Labs, an experimental platform for early Postgres tooling built around Lakebase Postgres, beginning with an Upgrade Assessment and Migration Assistant designed to help users prepare for major-version migrations. The read-only assessment analyzes a Neon project against version-specific rules to identify blockers, warnings, extension compatibility, schema issues, and logical-replication risks, while the assistant recommends migration approaches based on database size and downtime needs, including managed imports, pg_dump/pg_restore, and logical replication. For logical replication, the tool automates or monitors tasks such as schema setup, publication and subscription configuration, replication lag, sequence alignment, row-count checks, ANALYZE, rollback, and teardown, though users must still pause writes and change application connection settings during cutover. Neon emphasizes that the tools are experimental and should only be used with development or test workloads for now, while production users should follow established upgrade paths. The initiative addresses the complexity of Postgres upgrades, particularly as Postgres 19 approaches and Postgres 14 nears end of support, since Neon projects are tied to their original major version and require migration to a newly created target project. Neon Labs is built with Next.js, TypeScript, the Neon API, and direct catalog queries, does not store assessment data, separates assessments from migration actions, and is available both as a hosted service and an open-source local deployment.

Neon has renamed its CLI package and documentation command from neonctl to neon while retaining backward compatibility, and has added tools aimed at agent-driven and branch-first Postgres development. New onboarding commands configure authentication, MCP support, editor integrations, and agent skills, while link, checkout, and env pull establish local project context, select database branches, and synchronize branch-specific environment variables. The release also introduces database diagnostics through inspect db, schema comparison through diff, terminal-based snapshots and restores, direct Console access, local branch status, scoped API keys, and multi-account profiles. Additional capabilities include a general API fallback for automation, terminal management of Neon Auth and the Data API, a built-in psql fallback, expanded project creation options, and declarative configuration support, consolidating database, backend service, and automation workflows within the CLI.

Lakebase Postgres proposes a “lakebase” architecture that treats PostgreSQL’s write-ahead log (WAL), rather than mutable database files, as the durable and addressable source of truth, storing historical data in object storage while separating compute from storage. Standard Postgres compute nodes continue to execute SQL, manage transactions, and use RAM and local NVMe caches, while safekeepers replicate WAL records through quorum-based consensus and pageservers reconstruct page versions for a requested log sequence number (LSN) from base images and WAL deltas. This design aims to make commits independent of object-storage latency and enables historical reads, inexpensive point-in-time restoration, branching, and rapid read-replica creation because database states can be referenced by LSN instead of copied. The system uses immutable image and delta layers, along with persistent search trees that allow current and historical page lookups at similar cost. The article argues that storing durable transactional history in object storage can also support shared transactional and analytical access through columnar representations, reducing the need for separate replicated data pipelines. Lakebase Postgres is presented as available through Neon and Databricks, with integrations tailored to each platform.

Neon AI Gateway serves Databricks-hosted open-weight models through Databricks Foundation Model APIs, using an inference stack designed to improve real-world latency and throughput for workloads such as coding agents and multi-step tool loops. The post attributes performance to continuous batching, paged KV-cache allocation, combined prompt-prefill and token-decoding scheduling, TensorRT-LLM-based runtimes with custom GPU kernels, quantization, and hardware-specific multi-GPU tuning, particularly for Mixture-of-Experts models. It emphasizes prompt caching as especially useful for agents that repeatedly send stable system prompts, tool definitions, and examples, citing a Databricks batch-inference deployment of gpt-oss where a roughly 30% cache-hit rate corresponded with 2.5 times higher per-replica input-token throughput and three times lower median latency. Developers can access the gateway through compatible existing SDKs, pay standard model token rates without a Neon surcharge, and, during the beta period, use free tokens on eligible Neon plans in AWS us-east-2 to evaluate performance on their own workloads.

Neon has expanded backend observability by making beta logs for Neon Functions and Object Storage available outside its Console through the branch-scoped `neon logs` CLI, the Neon MCP server, Logs API, `@neon/sdk`, and raw Loki endpoints. Users can query logs by time range and filter them by source, service or scope, severity, message content, trace ID, or custom LogQL expressions, while CLI commands can also reveal available fields and their values and produce table, JSON, or YAML output. Neon Functions provide serverless Node.js compute alongside Lakebase Postgres, and Neon Object Storage provides S3-compatible storage that branches with project data. The new tooling is intended to help developers investigate cross-service issues such as failed uploads and function errors from a common interface, while planned future coverage includes Postgres compute, Managed Better Auth, and AI Gateway logs.

Neon has introduced AI Gateway in beta, positioning large language model access as a core backend service alongside its Lakebase Postgres database, Object Storage, Functions, and Managed Better Auth. Hosted on Databricks Foundation Model APIs infrastructure, the gateway provides a single Neon credential and OpenAI-compatible endpoint for accessing a catalog of proprietary and open-weight models, while retaining native provider routes for features such as OpenAI Responses, Gemini, and Anthropic Messages. Neon says usage is consolidated onto its invoice at providers’ published token prices without markup, with additional observability and spend controls planned. The service extends Neon’s branch-based development model by giving each database branch its own gateway host and lineage-scoped credentials, allowing preview, feature, and CI environments to isolate model calls alongside their data, files, functions, and authentication. Developers can enable the gateway through Neon configuration or the console and adapt existing SDK integrations primarily by replacing the API key and base URL; the company also offers templates for Vercel AI SDK and Mastra applications.

Vercel’s CLI now installs Marketplace providers’ agent skills alongside integrations, enabling users to provision a Neon Postgres database, connect it to a Vercel project, inject environment variables, and provide coding agents with Neon-specific guidance through a single command. The command supports interactive or flag-based setup, with Vercel-managed accounts billed through Vercel or Neon-managed accounts retaining Neon billing, and additional CLI commands for discovering, configuring, and opening integrations. When Preview is enabled, each Vercel Preview Deployment receives an isolated Neon database branch that shares storage until changes occur, allowing pull requests to use realistic but short-lived backend environments. Cleanup behavior differs between Vercel-managed and Neon-managed integrations, while preview deployments can also receive authentication configuration for Managed Better Auth. Neon skills can alternatively be installed directly, and users can access Neon tooling through the Vercel CLI, Cursor and Claude Code plugins, or Neon’s initialization command depending on whether they need provisioning, live database access, or complete project setup.

Neon for Herdr is a plugin that integrates Neon’s PostgreSQL platform into Herdr, a terminal workspace manager, providing a pane-based dashboard for managing projects, branches, computes, connection strings, databases, roles, and recent operations without leaving an editor or shell. Built with Ink, Effect, and the Neon TypeScript SDK, it supports fuzzy project search and keyboard-driven actions to create, rename, delete, reset, and inspect branches, as well as start, suspend, restart, and configure computes. Installation requires Herdr 0.7.0+, Node 22+, and macOS or Linux, while authentication uses Neon OAuth with PKCE and locally stored rotating credentials. Users can configure pane placement, refresh intervals, organization selection, and pooled versus direct connection strings, then assign their own Herdr keybinding because plugins cannot define one directly. The post highlights using the plugin to create a production-like branch for testing a migration, copy its connection string, reset failed attempts while retaining timestamped snapshots, and remove or suspend resources after testing is complete.

Neon Functions is a beta serverless compute offering designed to run Node.js 24 backend logic alongside a Neon branch and its Lakebase Postgres database, with database, Object Storage, and AI Gateway credentials injected automatically. The service targets data-adjacent workloads such as APIs, AI agents, WebSockets, and server-sent events by supporting longer-running requests and reusable PostgreSQL connection pools, while recommending that frontend hosting remain on platforms such as Vercel or Netlify. Functions are branch-scoped, so creating a database branch also creates isolated copies of deployed functions and associated backend resources, enabling preview environments that mirror production data structures without sharing data. Developers can declare functions through Neon’s backend-as-code neon.ts configuration, deploy HTTP handlers built with frameworks such as Hono, and use examples for REST APIs, agents, real-time applications, and bots. Neon notes that functions are not background job runners, although branch-scoped schedules and event triggers for storage, authentication, database, and platform events are planned, and the product is free during its beta period.

Databricks has announced that Electric, creator of the PGlite embedded WebAssembly Postgres database and the Electric real-time synchronization engine, is joining the Neon team within Databricks. Neon describes PGlite as a lightweight version of Postgres capable of running in browser tabs, sandboxes, embedded applications, and serverless environments, with weekly downloads reportedly growing from 1 million to 13 million over the past year. Electric’s sync technology continuously connects central Postgres databases with clients such as mobile apps, browsers, and agent sandboxes, addressing challenges including conflict resolution, partial replication, and reconnection. Neon says these capabilities will complement its existing Lakebase Postgres, authentication, storage, functions, and AI Gateway offerings as it develops backend infrastructure for agent-driven applications. The companies also note a shared technical history, as PGlite began as a WebAssembly Postgres experiment by Neon co-founder Stas Kelvich before Electric developed it into its current product.

Neon has launched Object Storage in beta, adding an S3-compatible, branch-aware file storage service to its backend platform alongside Postgres. Creating a Neon branch now forks both database data and object storage under one branch ID, giving child environments point-in-time access to inherited buckets and files while keeping later uploads, deletions, and modifications isolated through copy-on-write and versioned lineage mechanisms. The service is designed to make preview environments, pull request deployments, experiments, and agent workflows inexpensive and disposable, since no database pages or objects are duplicated until a branch diverges and compute scales to zero when idle. Object Storage supports standard S3 clients such as AWS SDKs, boto3, and the AWS CLI, while bucket access settings are managed through Neon rather than S3 ACLs or policies. Users can provision buckets through Neon’s neon.ts infrastructure-as-code configuration and deploy them with related backend services, including Functions, Auth, and AI Gateway, as Neon expands its broader concept of fully branchable backends.

Castform and Neon describe a workflow for reinforcement-learning post-training of small open-source models to perform agentic search over company data, claiming that a 4B model trained this way matched GPT-5.6 Sol’s retrieval accuracy at roughly one-hundredth the cost. The approach responds to a shift from one-shot embedding-based RAG systems toward multi-step agentic retrieval, where frontier-model search loops can add substantial latency and expense. Castform generates synthetic questions and ground-truth answers from documents stored in Neon Postgres, trains models to use Lakebase Search’s text, vector, and hybrid retrieval tools, and evaluates them through rewards for finding correct sources, citing them, and answering accurately. The platform also provides training-run observability to inspect reward progress and diagnose issues such as faulty tools or reward hacking. Neon’s autoscaling database infrastructure is presented as useful for the bursty workloads created by many parallel search rollouts, while its branching and time-travel capabilities could support isolated, resettable environments for training agents that modify data as well as retrieve it.

Neon has introduced a new permission structure featuring org-level roles and per-project permissions to enhance access control for agents and collaborators. This update addresses the limitations of the previous system where a single role was applied uniformly across all projects in an organization, which was suitable for small teams but inadequate for managing databases at scale. The new model allows for more granular access by combining baseline org-level roles—Admin, Editor, Viewer, and Collaborator—with project-specific permissions that can extend access on particular projects. This dual-layered approach ensures that agents can manage projects effectively, reducing the risk of over-permissioned keys and improving safety management. This structure also facilitates developer collaboration by enabling specific project access without exposing the entire organization's resources. While current Neon users will retain their existing access levels, the traditional project collaboration feature will be phased out in favor of the new system, which will soon be available for Vercel Managed Integration users.