ObsessionDB gives AI agents their own lane through company data

ObsessionDB wants AI agents to interrogate company data without slowing the applications that depend on it. The startup has launched a real-time warehouse built on its own fork of ClickHouse, with separate computing resources for agents and a shared underlying store of data.

The practical pitch is that a customer can turn a stream of agent transcripts, logs and traces into something queryable as it arrives. Those records can support analytics, but also APIs that feed a live product. That puts the warehouse closer to operational software than a dashboard someone checks at the end of the day.

In its launch announcement, ObsessionDB says it is already processing millions of real-time API requests for customers. Its largest customer moved from Elasticsearch and now serves APIs over 50 terabytes of data while ingesting hundreds of thousands of events a minute, according to the company. It did not identify that customer in the announcement.

The architecture grew out of the founders’ previous work on blockchain data infrastructure. Rafael Aviles, Marc Höffl and Marc Leuthardt say they had been running workloads at petabyte scale and repeatedly encountered limits on cost and speed. They took ClickHouse’s open-source engine and separated its computing layer from storage, keeping one copy of the data in object storage with stateless compute above it.

Watch ObsessionDB’s original launch demonstration.

That separation matters when the same dataset serves very different users. A production application needs predictable responses, while an agent may issue many exploratory queries before it finds an answer. ObsessionDB’s website describes isolated, read-only compute for agents over the same data, intended to keep that exploration from competing with production traffic.

The company also says it logs agent queries so customers can inspect what was read. It charges for compute hours rather than individual queries, a model that fits the repeated probing it expects from automated users. Those are specific operating choices; the broader claim that the system is cheaper than competing warehouses still depends on the workload.

ObsessionDB markets a cache mesh that keeps frequently accessed data in the query layer, with object storage used for durability. Its website lists a per-agent sandbox as early access. That distinction is useful: isolating read-only queries is available in its product description, while giving every agent run a disposable branch of production data remains a more limited offering.

The company is part of Y Combinator’s Fall 2026 batch. Its founders bring experience from Numia and other infrastructure businesses; Höffl also previously founded Laserfocus, a Winter 2020 YC company. That earlier accelerator history belongs to a different business.

For teams considering a move, ObsessionDB offers migration help, including benchmarking a customer’s workload and shadowing production traffic before cutover. That is where its cost-and-speed argument can become concrete: the queries already running in a customer’s system, measured against the bill and response times they have today.

‍