AI agents get deeper into the database as Postgres platforms open up their tooling

Agents move closer to the metal

PlanetScale spent the week widening what an AI assistant can see and do against a database. The MCP server can now pull branch metrics — CPU, memory, disk, replication lag, and connections for Vitess, Postgres, and Neki branches — so an agent debugging a slow query can check the last hour of utilization without leaving the editor. With that much reach, the organization-level cap on MCP query access arrives at the right moment: admins can set clients to no access, read-only, or read and write, applying to newly authorized clients while existing ones keep their grants.

Neon took the opposite tack on access, opening every model in its AI Gateway catalog to paid plans with prepaid credits — no access request required, Claude Opus 5.5 included. Neon also added a way to complain about it: a send_feedback MCP tool and a neon feedback CLI command, plus -y flags on neon deploy and neon config apply for unattended runs.

PlanetScale makes Postgres and migrations more elastic

PlanetScale Postgres gained independently configurable dedicated read replicas, creatable in the primary region or a second region from the same cloud provider with their own cluster size, replica count, storage, and Postgres parameters — manageable from the dashboard, API, CLI, or Terraform.

Vitess users got a smaller but sharper fix: deploy requests can retry only the failed tables of a partially failed migration instead of restarting the whole deploy, leaving completed tables at their existing progress. The PlanetScale CLI folded that in alongside Postgres extension toggles, VTGate parameter commands, human-readable sizes for keyspace --max-storage and --storage, and clearer messaging for expired or denied device logins.

Prisma's v8 candidate stops dropping data quietly

Three release candidates landed for Prisma 8, and the headline change is about safety: prisma migration plan and prisma db update now ask about every model or field a plan would drop, answered with --delete or the new --rename, which keeps the rows. The ORM client picked up row locks, and apply and scope became with and fragment. Earlier in the week, Prisma fixed a Bun install that could resolve two ArkType copies and break prisma orm init on a valid contract, and moved SQL contracts to explicit column data types like pg/text, with codecs validating every stored value and custom collection classes keeping their methods through the chain.

Tracing, scopes, and smaller cleanups

Supabase's CLI shipped a large batch of beta and next-channel iterations plus a stable v2.120.0 with opt-in OTLP tracing and persisted stack log history, the latter browsable via stack logs history and shipped to Analytics. The Client SDK patched a remote code execution vulnerability in its realtime-js example, corrected multipart field ordering for Storage uploads, and taught PostgREST to treat a TimeoutError as an abort in its retry guard; a canary also stopped postgrest-typegen from emitting computed fields in select('*').

Elsewhere, Turso's CLI added scopes to org-wide API tokens, and the Turso database pre-release continued its 0.8.3 line. Prisma's language tools cycled through a run of insider and stable builds, and roughly a half-dozen lower-priority releases rounded out the week without changing the story.