Installation
Prerequisites
- Node.js 20 or later, on macOS or Linux
- Claude Code or Codex, signed in
Claude Code
- Run this one command and you're set:
npx @fewr/cli
- Start a new chat. Fewr loads when a chat starts, so chats already open keep running without it.
- Type this in any chat to see your savings:
/fewr
Codex
- Run the same command:
npx @fewr/cli
- In Codex, type
/hooks. Find Fewr: SessionStart and Fewr: record usage, and trust both. Codex keeps new or changed hooks off until you do, so repeat this after each update.
- Start a new chat.
- Type this in any chat to see your savings:
$fewr
To remove Fewr: npx @fewr/cli uninstall.
Benchmark
Same task, same model, same effort, with and without Fewr. Each result is the average of two runs, and hidden checks graded every build.
Before / after
Same game, same model, same effort. Without Fewr: 6.3M tokens, $6.14, 37 minutes. With Fewr: 2.7M tokens, $3.36, 18 minutes. Both pass all 12 hidden checks and play the same. This is one of the two runs in the benchmark; across both, cost fell 43% on average.
Why Fewr is different
Most ways to save usage give up some quality to get there:
| Approach | What it does | The catch |
|---|
| A cheaper model | Runs the whole session on Sonnet or a mini model | The hard parts (architecture, design, security) get the weaker model too |
| Lower effort | Thinks less on every turn | Less thinking on exactly the turns that need it |
| Trimming and compression | Shortens prompts, context or replies | Mostly cuts cached context, which is already cheap. Saved 0–11% in our tests, and rules about how to work made models think more, not less |
| Proxies and routers | Re-route your subscription's traffic | Breaks Anthropic's and OpenAI's terms and puts your account at risk |
Fewer tokens isn't less usage. In a coding session, most tokens are the same context re-read from cache on every turn, and a cached token costs 5–10% of a fresh one. Cutting them shrinks the token count much more than the bill. What really costs is which model does the work: Haiku and Luna cost 40–100× less per token than Opus and GPT-6 Astra.
Fewr saves there. It doesn't turn anything down across the board; it splits the work:
- Your model keeps the decisions. Architecture, design, security judgement and review stay on the model and effort you chose.
- Each subtask gets the right model. Lookups, tests and well-defined code go to the cheapest model and effort that does them well, and escalate if they fail.
- Nothing in between. Fewr runs through Claude Code's and Codex's own hooks, with your login, on your machine.
- Measured, not promised. Every benchmark compares the same model at the same effort, with hidden checks grading both builds.
/fewr shows what it saves in your own chats.
Follow along
License
AGPL-3.0. Free to use, modify and share. If you distribute a modified version, or offer one as a network service, you must publish its source under the same license. The Fewr name, logo and website content are not covered; see NOTICE. © 2026 Florian Jousselin.
