Distill

Distill is an open-source coding agent harness and TUI published by Samuel Fajreldines (samuelfaj) under the Apache 2.0 license. It works with Grok and Codex subscriptions and with any model that speaks the OpenAI-compatible protocol. The latest listed release is v2.0.49.

What It Is

Distill is a terminal-based coding agent environment aimed at reducing token consumption while completing coding tasks. Users install it with a script on Mac, Linux or Windows, sign in with OpenRouter and/or a subscription, and run sessions from its TUI. The project states that its benchmarks show the same quality at up to 87% lower cost and up to 76% less time versus several native model baselines; these are the author's claims, documented in a separate BENCHMARKS file.

Model tiers and delegation

Distill organizes models into three tiers. The main model is required and owns each session, planning, specifying, delegating and reviewing work. An optional worker model runs delegated assignments handed over as precise specs. A utility model handles bounded tasks such as extraction, summaries and compression. Each tier accepts an effort level or auto, and tiers can be set from the Model tiers screen or with commands such as /model, /worker-model and /utility-model.

Always-on behavior

Two rule sets are built into every system prompt for the main model and its subagents. Caveman makes the agent write terse prose while keeping code, paths and error strings exact, with adjustable levels (lite, full, ultra). Ponytail has the agent ask whether code needs to exist and reuse existing code, the standard library or installed dependencies before writing the minimum that works, without cutting validation, error handling or security. The README also links docs on routing, local models, and the UltraCode, UltraFast and UltraSave execution modes.