The first thing worth knowing about Rudel — before the dashboards, before the dataset — is what it uploads. When a Claude Code session ends and the hook fires, the full transcript goes to ClickHouse. That means source code. It means tool output. It means whatever secrets happened to pass through the context window. The project documentation says so plainly, and it's the right place to start.
With that out of the way: Rudel is a genuinely useful piece of open-source infrastructure for anyone running Claude Code regularly. The platform, free at rudel.ai and fully self-hostable, gives developers a dashboard covering token spend, session duration, activity patterns, model selection, and sub-agent spawning. It's built on 1,573 real sessions that the team — operating under the ObsessionDB GitHub organization — collected before this week's launch.
Installation is straightforward. An npm package registers a Claude Code hook on the SessionEnd lifecycle event, a documented extension point in the official Hooks system. When a session concludes, the hook ships the transcript — including git context, project metadata, and sub-agent records — upstream for processing.
The sub-agent data is one of the more valuable outputs. Claude Code can delegate mid-task to parallel agent processes, and tracking how often that happens, and in what contexts, matters for anyone watching costs or trying to understand why a session ran long. Rudel surfaces it alongside the standard token burn metrics.
The 1,573-session foundation distinguishes this launch from most agent tooling releases, which tend to ship with toy benchmarks or no usage data at all. Aggregate session duration distributions, sub-agent invocation rates, and token consumption trends from real developer sessions give the dashboards a layer of context that otherwise takes months to accumulate. The ClickHouse backend puts Rudel in technical company with production-grade LLM observability tools like Langfuse and LangSmith.
Back to the privacy question, which deserves more than a footnote: for personal projects and sandbox environments, the upload architecture is probably acceptable. For anything touching customer data, production systems, or credentials that have not been rotated since your last Claude Code session, read the documentation carefully before installing. The project is MIT-licensed and built with Bun and TypeScript.