We're building the database we always wished we had.
Antfly is a distributed document database for the AI era — proven distributed-systems engineering and first-class AI in a single binary you can actually run.
We're builders, platform engineers, and data scientists. We spent years shipping search, retrieval, and ML systems, so when teams started putting agents into production, the gap was familiar to us.
Search used to be something people did: you ran a query, read the results, and decided what mattered. Agents don't work that way. They retrieve and act, with no one in the loop to notice a stale policy or a missed clause. What an agent retrieves is what it acts on. That makes retrieval a correctness layer, and most of the infrastructure underneath it was built for a different job.
The usual approach is to assemble it yourself — a vector store, an embedding API, a reranker, a graph database, and the code to hold them together. We did this more than once. It's a lot to run, and at the end you still don't have a system that learns; the pieces were designed separately and stay that way. What we wanted was one engine where storage, retrieval, ranking, and inference share a feedback loop, so it gets sharper as it's used.
That's Antfly — the database we wanted, but it didn't exist. So we're building it.
- Operational simplicity
- One binary. If it's hard to run, nothing else matters.
- Distributed by design
- One node to many, with no rewrite in between.
- AI is the default
- Hybrid search, embeddings, and reranking belong in the database — not in a layer of services around it.
- Local-first
- Your data and your models never have to leave your own hardware.
- Boring where it counts
- Proven distributed-systems engineering underneath the novel parts.
- Built for builders
- A clean API and SDKs that get out of your way.
The team behind the Ant
Engineers, researchers, and builders writing the thoughts and shipping Antfly.
