ObsessionDB is a fully managed ClickHouse service for teams whose agents ingest large volumes of data and query it far more often than a human analyst. Compute and storage are separated. Durable data sits on object storage, compute nodes are stateless, and a distributed NVMe cache keeps hot data on local disk so reads do not wait on remote storage. The launch is aimed at a cost problem the founders say they already hit. Warehouses were priced for people running a handful of queries, while agents ask questions continuously, including over raw history that teams often drop to keep the bill down. It keeps ClickHouse SQL, functions, drivers, and the wire protocol, and replaces the MergeTree storage family with SharedMergeTree. Metadata lives in a coordination layer, writes persist to object storage, and a cache mesh serves queries. Hosting runs on AWS, Google Cloud, and Azure, in every public region, with a free developer tier up to 4 vCPU and 8 GB of RAM. The bill is tied to the compute and storage you provision, not to bytes scanned or questions asked. Rafael Aviles, Marc Höffl, and Marc Leuthardt are the founders. They previously built Numia, where they ran real-time data workloads at petabyte scale, and ObsessionDB is in Y Combinator's Fall 2026 batch.