Cloudflare details Town Lake data platform and its Skipper AI agent

Cloudflare described its internal Town Lake analytics platform and Skipper data agent on May 28. The engineering post explains how Town Lake puts one governed SQL interface over systems including Postgres, ClickHouse, BigQuery, Kafka-derived data, and Iceberg tables stored in R2.

A federated interface with governance in the path

Town Lake uses Trino as its query engine, allowing one query to join data from different stores without first copying everything into a single database. Cloudflare distinguishes fresh, unsampled sources needed for work such as billing or security investigations from downsampled sources suited to exploration. Its supporting services catalog data, identify sensitive fields, manage time-bounded access, and record audit history.

Skipper sits above that layer. It finds likely datasets, inspects schema and lineage, writes SQL, runs the query, and returns a table or chart. Cloudflare also exposes Skipper tools through an MCP server using a code-mode design: the model searches the tool surface and executes a JavaScript workflow instead of receiving every tool definition in the prompt.

The hard problem remains trustworthy context

A natural-language interface makes access easier, but correctness depends on catalog quality and permissions. Teams building a similar agent should make the generated SQL visible, link results to the source tables and time range, and refuse ambiguous joins rather than silently guessing.

Evaluation should include questions with stale tables, similarly named metrics, restricted columns, and empty results. The agent must preserve the caller’s authorization throughout discovery and execution; it should not gain broader data access because its interface is conversational.

Town Lake and Skipper show how a data platform can give an agent a controlled query surface. Their value comes from the catalog, lineage, access checks, and auditability underneath the chat box.