Argon is Postgres whose tables are also live Apache Iceberg tables — and whose built-in AI interface can tell an agent exactly what changed since it last looked. Two things no other database does. Watch both, live, below.
While this demo sleeps, the real music industry keeps editing MusicBrainz. Waking it applies every real change since the last visitor — that catch-up is the first act.
Argon is a Postgres database that keeps an always-current analytical copy of itself — with no pipeline — and whose built-in AI interface can tell an assistant exactly what changed since it last looked. Databases normally need a second system (and a team) to do the first, and nothing on the market does the second: staleness — "what did I miss?" — is a question Argon answers natively, because every change already flows through it.
That's why the demo starts asleep. Waking it up is the product: hours of real music-industry edits become queryable, versioned, AI-visible analytics in seconds, untouched by human hands.
Argon tails the Postgres WAL into Apache Iceberg (merge-on-read row-deltas with equality deletes, multi-table atomic commits through the REST catalog, per-table changelog sidecars carrying LSN/timestamp/op). The MCP layer serves LSN-pinned, token-budget- enforced tools from the lake tier — the OLTP primary never serves analytical reads. Every claim in the act cards maps to a mechanism you can inspect with the "under the hood" toggle; the storage engine continues the Neon lineage (separated compute/storage, copy-on-write branches, crash-only ingest that replays from its own watermark).
It also ships as a Helm chart with per-cloud presets — verified on kind with pod-loss writes, piloted on AKS against real Azure Blob/ADLS. This demo deliberately runs on a single VM: Kubernetes is a deployment target, not a requirement.
Coming next: retrieval — the changelog you can watch below already computes exactly which rows an embedding index would need to refresh.
Live data: MusicBrainz (via the MetaBrainz Live Data Feed; we are not affiliated with MetaBrainz). Database content CC0/CC-BY-NC-SA per MusicBrainz licensing.
claude mcp add --transport http argon \ https://demo.argondb.ai/mcp
Nine read-only, token-budgeted tools against this same live catalog: pin a consistent view, query it, ask what changed.
ATTACH 'argon' AS lake (TYPE iceberg, ENDPOINT 'https://catalog.argondb.ai', AUTHORIZATION_TYPE 'none'); SELECT count(*) FROM lake.br_main.recording;
The same tables, read by a stock engine. Open Iceberg — no lock-in. Until the public catalog endpoint ships, Act 6 runs this exact read live, server-side.
Retrieval ("the database feeds the model") is next on the roadmap — the changelog you watched already computes exactly which rows would need re-embedding.