// the find
marmotdata/marmot
The open-source context layer for your AI. Catalog your tables, topics, queues and APIs then expose real metadata to your AI agents.
Marmot is a data catalog written in Go that ships as a single binary with a built-in MCP server, so AI agents can query real metadata about your tables, topics, and APIs instead of hallucinating schemas. It targets teams who find DataHub or Amundsen overkill but still need lineage tracking and ownership management. Early-stage: 576 stars, 19 forks.
Single-binary deploy is genuine for dev and small teams — no separate search cluster, no vector store, just run it. The Helm chart exists if you need it but you're not forced into it on day one. MCP integration is first-class rather than bolted on — AI agents get certified metadata through a proper server, not a scrape-and-hope wrapper. Plugin breadth covers a real data platform: BigQuery, ClickHouse, Airflow, Confluent, dbt, AsyncAPI across 12 protocol variants, Azure Blob. OpenLineage support means you can feed lineage from existing dbt or Spark jobs without rewiring everything to Marmot's native format.
19 forks is very low for the claimed scope — most plugins have a source.go with no corresponding test file, so you are the integration test if you adopt a less-common connector in production. The 'single binary' story breaks down at scale: the production Helm chart pulls in CloudNativePG and a Kubernetes operator with CRDs, RBAC, and job templates, which is not simpler than alternatives it claims to beat. The operator's runs/scheduling system adds real complexity for teams who just want scheduled ingestion without a k8s footprint. All real documentation lives at an external site with no offline fallback — the README is a marketing FAQ table, and if marmotdata.io lags the code you're reading source.