Summary

  • RavenDB launched Quill on 8 September for PostgreSQL, SQL Server and MySQL; agents use an internal copy while the original SQL system remains authoritative.
  • The write meter includes source-data ingestion and changes, agent messages and embedding generation. Unmetered reads do not make the model bill disappear.

A quiet agent can sit beside a busy database. That distinction matters for Quill, the enterprise agent platform RavenDB announced on 8 September. If selected source records change frequently, keeping the agent's context current can consume write units even when employees ask relatively few questions. This is an implication of the published charging model, not a measurement of a customer's bill.

Quill's sales proposition is that a company need not migrate its SQL system of record to introduce agents. The launch names PostgreSQL, SQL Server and MySQL. “No migration” has a precise boundary here: the production database keeps its role, but the agent service still maintains another representation of its data.

The copy is part of the product

The technical overview describes a Docker service running beside the SQL database. Quill initially reads the selected tables into its own internal database. Subsequent inserts, updates and deletions reach that copy through change data capture, or CDC. Agents query the mirror rather than replacing the application's existing SQL traffic.

That arrangement separates an agent's retrieval workload from the application's normal query path. It does not establish zero operating burden. The initial copy can take time for a large database, although the documentation says an interrupted copy can resume. Administrators may also need to enable change recording or provide the necessary permissions. Continuous synchronisation is not evidence of zero lag under every workload.

The commercial significance becomes clearer on the pricing page. Quill uses Write Request Units, or WRUs. Source-record ingestion, later changes and deletions, messages sent to an agent and embedding generation all count as writes. The measure is per document rather than per byte; an order and its line items are separate writes. Reads, queries and internal index maintenance are not metered by Quill.

What the launch prices do—and do not—price

Starter is listed at $499 a month with one million WRUs included and an additional rate of $0.60 per thousand. Pro is $1,799 with eight million included and $0.30 per additional thousand. These are expressly launch prices, and the page says included volumes may be adjusted.

There is a contracting detail worth settling before turning those rates into an invoice forecast: the FAQ also says WRUs are charged in million-unit increments. A displayed rate per thousand does not, by itself, resolve the bill's rounding or commitment granularity.

The 90-day full-featured trial includes unlimited WRUs. Initial ingestion still consumes units; the trial covers them. Its absence of a WRU charge is therefore not evidence that initial loading or subsequent replication has no economic weight after the trial.

Model tokens are outside the quoted subscription and are paid to the model provider separately. A self-hosted model may avoid that provider's token charge, but not the hardware, energy or operating work required to run it. Similarly, free reads within Quill's meter do not mean an all-in free inference request.

The buyer's useful starting point is consequently the selected data's rate of change, alongside messages and embeddings—not simply headcount or query volume. Public documentation does not establish a customer's steady-state WRUs, measured replication lag or all retry and resynchronisation charging details. Quill offers a route around moving the authoritative SQL system. The evidence does not support treating the context store that replaces that migration project as maintenance-free.