Developer zdk released lowfat, a lightweight CLI tool that trims unnecessary command output to reduce AI agent token consumption. The single binary filters verbose dumps like kubectl get -o yaml before they reach agents, preventing expensive token waste on irrelevant data.

lowfat operates as a shell wrapper or agent hook with a plugin architecture for custom filtering. Its design philosophy emphasizes a small core that users can extend rather than bundling every possible command. The tool achieves 91.8% token savings across common operations like kubectl, docker, and git commands, based on two months of personal usage data.

Unlike alternatives that bundle comprehensive filter sets, lowfat lets teams customize filters for their specific toolchains including internal enterprise CLIs. The local-first approach means zero telemetry and user-owned data. Aggressive filtering levels can be dialed down to preserve information agents might need.

Installation works via cargo or brew, with integrations for Claude Code, OpenCode, and Pi agent. Users prefix commands with lowfat or configure automatic filtering. The Apache-2.0 licensed project addresses a genuine pain point in AI-assisted development workflows.

GitHub: https://github.com/zdk/lowfat