I built Iris after noticing that a fresh Claude Code session was carrying far more context than I expected before I had done any meaningful work. In one setup, the baseline was 28.0k tokens, with 25.6k coming from system tools. After removing optional tools I wasn't using, that dropped to 5.6k. Iris shows where that context is going, which tool schemas are being carried, whether those tools are actually used, and how the request maps to API-equivalent cost and quota consumption. From there, you can remove unused schemas, set Allow / Ask / Deny rules for tool execution, and review the tool activity from the session in the Flight Recorder. Iris runs locally, is MIT licensed, and is available through npm. I'd like feedback from people using Claude Code heavily, especially around context usage, cost visibility, quota efficiency, and Guard behavior.

Hammad Abbasi

See Claude Code context. Control what it can do.

See Claude Code context. Control what it can do.