Loams

What it does

The first line of the README is the whole premise: don't rent a cloud. Loams keeps data in the user's own object-storage bucket, in open formats, and the engine speaks existing wire protocols — so a team can point a Qdrant, Elasticsearch, Postgres or MySQL client at it instead of adopting a new API.

Around that sits tooling for building software with AI agents: an agent daemon, durable execution, a console, a desktop app, a CLI and SDKs.

Why it exists

Most data platforms that serve AI workloads ask you to move your data into their account and learn their API. Loams bets on the opposite: your bucket, open formats, protocols you already have clients for. It's Apache-2.0 with no open core and no paid edition.

How it's built

A Rust monorepo of roughly eighty crates, grouped by concern — the engine (query, collections, an HNSW index, text search, the log and the store), a compatibility layer per wire protocol, metadata and storage, durable execution, the graph, and the agent daemon. It builds on established components rather than reinventing them: DataFusion for query, Lance for columnar storage, Tantivy for text, SlateDB, openraft and TiKV among others.

My part

I'm one of two listed maintainers in a flat governance structure. I'm deliberately not claiming authorship of specific subsystems here: the public history doesn't attribute individual components, and the design decisions belong to the project.

Evidence, as the project publishes it

Loams makes no headline benchmark claims. What it publishes is more useful: raw write-path and commit benchmarks comparing NVMe write strategies — p50/p99 latency, throughput, CPU per transaction, noise bands over repeated runs — including gate reports where configurations fail their target. Roadmap targets are stated as gates, not as results.

Where it stands

Version 0.1.2, a patch release on the first release line, shipped on 26 October 2026. The project says plainly that APIs and on-disk formats may still change between minor versions.