Rudel, an MIT-licensed analytics dashboard for Claude Code, launched this week on GitHub under the obsessiondb organization and was featured on Hacker News as a Show HN submission. The tool gives developers and teams visibility into token usage, session duration, activity patterns, model usage, and <a href="/news/2026-03-14-axe-a-12mb-go-binary-for-unix-style-llm-agent-orchestration">sub-agent invocations</a>. Installation is straightforward: a CLI available via npm registers a Claude Code lifecycle hook tied to the SessionEnd event, which automatically uploads session transcripts to a ClickHouse database for processing and visualization. A hosted version is available at rudel.ai with free account creation, and self-hosting documentation is provided for teams with data-residency requirements.
The backend choice is worth unpacking. Rather than PostgreSQL or SQLite — the default reach for most developer tooling projects — Rudel uses ClickHouse, the OLAP columnar database originally built at Yandex and optimized for sub-second aggregation queries across billions of rows. The obsessiondb team is already publishing cross-user behavioral statistics drawn from their aggregated dataset, including the figure that 26% of Claude Code sessions are abandoned within the first 60 seconds. That kind of cohort-level query is exactly what ClickHouse handles efficiently and what would strain a row-oriented store. Whether the founders deliberately designed for fleet-scale from day one or grew into the architecture isn't stated, but the infrastructure suggests they're not building just for individual developers.
The Hacker News thread brought up an obvious competitive question: Claude Code already ships with a built-in /insights command. That positions Rudel as a persistent, historically queryable, team-oriented layer rather than a replacement for something that doesn't exist. Commenters also traded session management advice — short focused sessions, single-purpose task scoping, current AGENTS.md files for fast onboarding. One commenter noted, specifically about their own workflow, that LLM consistency is a real challenge and that Claude can drift from instructions in <a href="/news/2026-03-14-axe-a-12mb-go-binary-for-unix-style-llm-agent-orchestration">SKILL.md files</a> mid-session.
The privacy question is real. Rudel ingests full session transcripts, which can include source code, file contents, command output, and credentials. The project README flags this explicitly and advises users to enable uploads only on projects where they are comfortable sharing that data. For the hosted service, the team says it does not access personal data in uploaded transcripts, but directs users to review the full privacy policy before enabling team-wide uploads. Self-hosting lets organizations avoid sending raw transcript data to a third party entirely — and given the 26% early-abandon rate the team is already surfacing from aggregate data, that's not a hypothetical concern for every engineering org.