Slimming command output so LLMs stop reading the fine print

What it does lowfat is a local CLI filter that strips boilerplate from common command output—version-control status hints, container headers, directory permissions—before it reaches an AI coding agent. It bundles six curated built-ins and offers three compression levels that trade verbosity for token savings. The tool runs offline with no telemetry and plugs into several agent environments to intercept output automatically.

The interesting bit The project treats token savings as a user-owned optimization problem rather than a black-box service. It exposes a custom .lf DSL for writing your own filters, ranks your most-run commands with a local history feature so you know where to optimize, and even publishes head-to-head token counts against rival rtk with the honest caveat that the numbers are directional, not gospel. That transparency is rarer than it should be.

Key highlights

• Ships with six built-in filters for version control, containers, search, and directory listings.
• Extensible via a custom .lf filter DSL, plus shell and Python escape hatches.
• Integrates with Claude Code, OpenCode, Pi, and generic shell hooks to intercept output transparently.
• No telemetry; analytics like stats and history stay local.
• Explicitly compares itself to rtk as a minimal, user-extensible core versus a batteries-included suite.

Caveats

• The bundled token-savings percentages are measured on single-command output, not end-to-end agent context, so your actual mileage will vary.
• The head-to-head comparison with rtk is explicitly labeled as a single small run and directional, not a rigorous benchmark.
• Aggressive ultra filtering can drop data your agent actually needs; the README warns you to verify output before trusting it.

Verdict Worth a look if you pay for AI coding agents by the token and want fine-grained control over what context reaches the model. Skip it if you prefer an all-in-one solution like rtk that covers 100+ ecosystems out of the box.