alex shapalov (@alxshp) on X

What it claims

PGX is a Postgres runtime written in Rust that runs thousands of isolated Postgres databases in one process, aimed at AI agents. Its project site is pgx.dev. A companion agent skill gives every Claude Code agent its own Postgres database automatically, with no Docker, no cloud service and no manual database setup. The launch post states two different per-database footprints: "70 KB Postgres database per AI agent" in the skill announcement, and "~0.7 MB per idle database" in the runtime announcement. It also claims "up to 100× less idle memory" but does not name the comparison.

Evidence

The figures come from two X posts by alex shapalov (@alxshp), each with a video. The posts give no hardware, Postgres version, workload or baseline for any number. The runtime figures are 1,000 databases in 659 MB RAM, about 0.7 MB per idle database, and an 85 ms clone of a 674 MB database. The skill figures are a test with 200 concurrent agents and "zero database conflicts". The posts do not say what counted as a conflict.

How to use it

The skill installs with one command:

npx skills add pgrundev/pgx --skill pgx

The command implies the repository is pgrundev/pgx. The post describes the result as one agent, one database, with cloning available.

Limits

The 70 KB and 0.7 MB figures differ by about 10×, and the posts do not reconcile them. The 70 KB figure may describe on-disk size, but the post does not say so. The "100× less idle memory" claim has no stated baseline, such as one Postgres server per database. The posts call PGX "free" and "open source" but name no licence. They also do not say how much of Postgres PGX supports, such as which extensions, which wire protocol version, or how durability works. Those details are not in the posts and would need checking in the repository before adoption.