toonarrow
The open-source, multi-CLI alternative to Fermat's Last Token.
toonarrow is a local proxy that cuts the tokens your coding agents send, by rewriting their context before it
reaches the provider. It works with Claude Code, Codex, OpenCode and Kora (Desktop and CLI), runs entirely on your machine, needs
no account, and is Apache-2.0 licensed.
"I have discovered a truly marvelous proof of this, which this margin is too narrow to contain."
Pierre de Fermat, in the margin of his Arithmetica.
Your context window is the margin. toonarrow makes the proof fit.
Why
Coding agents resend the whole conversation on every turn: the system prompt, every tool definition, every file
they read and every command output. Most of it is noise the model does not need verbatim, and you pay for it again on
every request.
Measured on real sessions:
| What | Before | After |
|---|
| OpenCode system prompt with 688 installed skills (first request, gpt-6-luna) | 85,436 input tokens | 25,559 (-70%) |
| Codex skill listing with 519 skills (every session start, lossless) | 39,367 chars | 17,775 (-55%) |
| Output tokens of the main agent with the short-answer rule (Claude Code, Haiku, 2 paired runs) | 2,933 / 2,278 | 1,617 / 1,748 |
Shell outputs from a real TypeScript monorepo (git log, grep, tsc, ls -la, ...) | 100% | 56% |
Over a whole working history the shell rules matter less than the table suggests: most of what agents read is file
content (cat, sed -n, Read), which toonarrow passes through untouched. On the author's last 60 days
(toonarrow estimate, every tool result counted once) they removed 2.0% of Claude Code's tool-result tokens, 2.3% of
Codex's, 8.3% of OpenCode's and 4.2% of Kora's. The large, repeated savings come from the skill listings, the
short-answer rule and, on Claude Code, the file tools. Run toonarrow estimate to get the numbers for your own
sessions before routing any traffic through it.
What it does
- Compresses shell output. Per-command rules for
ls -l, git log/status, grep/rg, find/fd, pytest,
vitest/jest, tsc, go test, cargo, npm/pip install, make, applied only where the command that
produced the output is certain: a simple command, the last step of cd dir && …, or the head of a pipe whose
other stages only select lines (| head, | grep pattern). Codex's code-mode scripts are read the same way,
one result per command, including the output of a command still running (write_stdin). Every compressed
output tells the model how to get the raw one (| cat).
- Compacts skill listings. Codex's listing is regrouped by skill root without losing a path (always on).
With
"skills": true, the skills you actually use keep their description in every client's listing and the
rest are listed by name, still loading on demand; toonarrow skills --write ranks them from your local history.
- Replaces file tools. An MCP server with
Search (several files per call, folded grep results, tree-sitter
outlines) and Edit (atomic batched edits with a read-gate): on by default with toonarrow claude, opt-in with
--toonarrow-tools for OpenCode and Codex.
- Keeps answers short. A short-answer rule appended to the main agent's instructions (on Codex, after your own
developer_instructions, never instead of them).
- Keeps OpenAI's incremental turns. Codex, OpenCode and Kora talk to the OpenAI Responses API over a WebSocket
that sends only the new items of each turn; toonarrow relays it and rewrites each turn like an HTTP request.
- Shows what it saved. A local dashboard and
toonarrow stats, in tokens, per day, session, model and client;
toonarrow estimate replays your local history offline.
What each client gets
| Claude Code | Codex | OpenCode | Kora Desktop / CLI |
|---|
| Shell output rules | Bash | code-mode exec, exec_command, write_stdin | bash/shell | executeCommand |
| Search result rewrite | - | - | grep | searchCode |
| Skill listing | - | regrouped (lossless), descriptions with "skills": true | with "skills": true | with "skills": true |
File tools (Search, Edit) | on (toonarrow claude) | --toonarrow-tools | --toonarrow-tools | - |
| Short-answer rule | on | on | on | - |
| Transport | HTTP | Responses WebSocket or HTTP | HTTP, Responses WebSocket with an OpenAI sign-in | HTTP, Responses WebSocket with the OpenAI provider |
Guarantees
- Your prompt cache survives. Every rewrite is a pure, deterministic and idempotent function of the block it
touches: the same history produces the same bytes on every request.
- toonarrow is never a dependency. When it 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.
toonarrow and Fermat's Last Token
Fermat's Last Token showed that a local proxy plus better file tools can cut an agent's bill. toonarrow takes the
same idea to every coding agent, in the open.
| toonarrow | Fermat's Last Token |
|---|
| Source | Open source (Apache-2.0) | Proprietary |
| Account | None | Required |
| Clients | Claude Code, Codex, OpenCode, Kora Desktop and CLI | Claude Code |
| Dashboard | Local, in tokens | Cloud |
Comparison 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
Requirements: macOS or Linux (with a systemd user session), Node.js >= 24.
npm install -g toonarrow
toonarrow install
From source:
git clone https://github.com/andsigno82/toonarrow.git
cd toonarrow
npm install
node src/cli.ts install
This is the beginning of the README. Read the rest at the original. Read the full README ↗
