Your context window is the margin.

Measured on real sessions

Coding agents resend the whole conversation on every turn: system prompt, tool definitions, every file they read, every command output. Most of it the model does not need verbatim.

• OpenCode system prompt with 688 installed skills
  85,436 → 25,559 input tokens−70%
• Shell output from a TypeScript monorepo: git log, grep, tsc, ls -la
  100% → 56% of the original size−44%
• Output tokens of Claude Code with the short-answer rule
  2,606 → 1,683 tokens, average of 2 paired runs−35%

What it trims

Shell output

Per-command rules for ls -l, git log, grep, tsc, test runners, package installs and more. Pipes and && chains are left alone, and every trimmed output tells the model how to get the raw one.

Skill listings

Skills you actually use keep their description; the rest are listed by name and still load on demand. toonarrow skills --write picks them from your own history.

File tools

An MCP server with Search, which reads several files per call and folds grep results, and Edit, which applies batched edits atomically.

Long answers

A short-answer rule added to the main agent's system prompt.

Nothing else changes

Every rewrite is a pure, deterministic function of the block it touches, so the same history produces the same bytes on every request and your provider's prompt cache keeps hitting.

When toonarrow is not running, your agents run exactly as if it were not installed. Nothing leaves your machine except the requests your agents were already sending: no account, no telemetry.

It works with the coding agents you already use, started exactly as before:

• Claude Code
• Codex
• OpenCode
• Kora

Kora, on the desktop and in the terminal

Kora is a free, native AI workspace for macOS: chat, editor, terminal and Git in one window, with your own models and keys. toonarrow works with both Kora Desktop and Kora CLI, which share the same sessions.

Kora Desktop

Turn on Settings › Smart Token Engine › Context proxy. Before each request Kora checks that toonarrow is up, and goes straight to the provider when it is not, so the switch can stay on. It covers the Kora, ByteDance, Kimi Code, custom and ChatGPT sign-in providers.

Kora CLI

Nothing to set: keep typing kora. When toonarrow is running, the shim points Kora at it through KORA_LLM_PROXY, the way ANTHROPIC_BASE_URL works for Claude Code.

In both, the skill listing and the output of executeCommand are trimmed, and the dashboard shows Desktop and CLI as separate clients.

toonarrow and Fermat's Last Token

Fermat's Last Token showed that a local proxy and better file tools can cut an agent's bill. toonarrow takes the same idea to every coding agent, in the open.

Based on public information as of October 2026. toonarrow is not affiliated with or endorsed by Quotient Labs; Fermat's Last Token is their product and trademark.

Install

1. Install the package and the background service:

   npm install -g https://github.com/andsigno82/toonarrow/releases/download/v0.1.0/toonarrow-0.1.0.tgz
   toonarrow install

2. Open a new shell, so the shims come first in your PATH.
3. Use your agents as usual, then see what was saved with toonarrow dashboard.

toonarrow uninstall removes the service, the command and the shims; your usage data stays in ~/.toonarrow.