Documentation

Distill.codes helps AI coding agents keep changes lean by reducing token noise, repetitive editing, and unnecessary cycles.
Use this guide to set up, run, and measure your results.

Quick links

Overview

Distill.codes is for teams and developers using AI coding agents for regular coding tasks. If you want less output waste, more predictable PRs, and lower infrastructure cost while keeping your existing tools, this is for you.

Distill.codes does not replace your agent, provider, or credentials. You keep your current Claude / OpenAI access and only route through Distill.codes as an optimization layer.

How it works

Distill.codes sits as a lightweight layer between your agent and the model API. It helps the agent complete tasks with less noise by improving how prompts, context, and outputs are handled throughout the task cycle.

In practice, this means faster iterations and cleaner final patches when the task is non-trivial: fewer unnecessary changes, less repeated back-and-forth, and fewer extra lines to review.

Distill.codes is intentionally minimal: one config update, existing workflows, and your normal agent login/session model.

Setup

After signup, copy your Distill.codes proxy key from the dashboard and apply one config value for your chosen agent.

  1. Open the Distill.codes dashboard.
  2. Create/verify your proxy key.
  3. Copy the generated URL for your provider.
    • Use ANTHROPIC_BASE_URL for Claude Code in ~/.claude/settings.json.
    • Use openai_base_url for Codex in ~/.codex/config.toml.
  4. Restart the agent process and run your normal workflow.

Claude Code setup

File: ~/.claude/settings.json.

Add your Distill.codes URL into settings so Claude API requests go through the proxy.

{
  "theme": "light",
  "env": {
    "ANTHROPIC_BASE_URL": 
      "https://proxy.distill.codes/{token}/essential/anthropic"
  }
}

Replace {token} with the token from your Distill.codes dashboard.

In Claude Code, run /model and choose the newest Claude model with maximum effort. Recommended: Opus 4.8 or newer. Minimum: Sonnet 4.6 with Max effort.

Codex setup

File: ~/.codex/config.toml.

Put the Distill.codes URL in the config value:

openai_base_url = "https://proxy.distill.codes/{token}/essential/openai"

Replace {token} and restart Codex.

In Codex, run /model and choose the newest Codex model with the highest available reasoning. Recommended: GPT-5.6 with xhigh or higher. Minimum: GPT-5.5 with xhigh.

Test it

To validate the effect, run the same task twice:

Use separate fresh directories or worktrees so runs do not share state.

Compare the two results on output tokens, wall-clock time, operations count, changed files, added/removed LOC, and final behavior.

Keep model, prompt, repository, and task definition as identical as possible.

Benchmarks

Use paired tasks to compare Distill.codes and direct provider runs.

Focus on the same repo, same task, same model, and same prompt. Track output tokens, wall-clock time, operations count, and changed LOC.

View benchmark notes and public results

Best practices

FAQ

Will Distill.codes always reduce token usage?

Not always. Some short or already compact tasks may show little difference.

Do you store prompts or generated source code?

No. Distill.codes stores no prompts, code, or full request/response bodies.

Can I switch back to direct provider access?

Yes. Remove the Distill.codes line from your local config and restart the agent.

Who is this for?

Anyone using Claude Code or Codex for real software work and wanting cleaner output economics.

Feedback

Use the public feedback repository for bug reports, feature requests, and product feedback.

Do not post API keys, proxy URLs, billing details, private source code, prompts, logs, or customer data in public issues.