A 1.7B-parameter shrink ray for CLI log spam

What it does

distill is a CLI filter that sits between your terminal and your LLM. You pipe in noisy command output—logs, search results, test failures—and it returns a drastically shorter summary that preserves the information the LLM actually needs. The README claims it can cut token counts by up to 99 percent, and the included example drops a 7,648-token search dump to 99 tokens.

The interesting bit

Instead of relying on a cloud API, it runs a custom 1.7B-parameter model locally via MLX, which means your data never leaves the machine and you don’t need a GPU farm. The project also pushes a “Distill Language”—a compressed syntax it claims helps LLMs think more efficiently, though the README treats this more as a teaser than a specification.

Key highlights

• Runs fully offline on a local 1.7B-parameter, 4-bit quantized model.
• Targets the specific problem of token bloat from agent and CLI tool outputs.
• Ships with a concrete before/after example showing a ~98.7% token reduction.
• Pitched at users of claude-code, Codex, and other agentic workflows.
• Requires roughly 8 GB of RAM minimum; 16 GB for comfortable use.

Caveats

• The “up to 99%” savings claim comes from the project’s own single example; your mileage will vary depending on output density.
• The “Distill Language” concept is illustrated with a screenshot but lacks technical detail on how it integrates with downstream LLMs.
• The README opens with an ad for a separate commercial client (remotecode.io), which slightly muddles whether the open-source tool is the main event or the funnel.

Verdict

Worth a look if you’re burning through context windows with build logs or search floods and want a local, privacy-friendly pre-processor. Skip it if you need guaranteed structured output or a well-documented integration API; this is early-stage plumbing.