Agent Skill · $15 one-time
TokenDiet
Cut the tokens your coding agent spends on every request, without losing quality. TokenDiet measures your instruction files, finds what can go, and trims it with your approval.
Works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf and Cline.
Why your requests are heavier than they need to be
Every request carries your CLAUDE.md, AGENTS.md, rules files and skill descriptions. Over time they fill up with repeated rules, release checklists, code samples and old history, and you pay for all of it on every turn, in tokens and in usage limits.
Lossless by design
What TokenDiet does
- Removes repetition and noise
- Moves instructions so they load only when needed
- Asks before every change
- Backs up every file it edits
- Measures again so you see the result
What it never does on its own
- Lower your model, effort or thinking
- Make replies terser
- Delete rules your agent still needs
- Disable your MCP servers or tools
When one of these would save tokens, TokenDiet explains the trade-off and leaves the choice to you.
What it finds
Repeated rules
The same instructions in CLAUDE.md, AGENTS.md and rules files, or twice in one file.
Workflow sections
Release, review or migration steps that load on every request but matter only sometimes. TokenDiet moves them into on-demand skills or scoped rules.
History
Change logs and past decisions that rarely help with today's work.
Code samples
Long examples inside instruction files, replaced with a pointer to a real file.
Long files
Files past 200 lines, the length Anthropic recommends staying under for CLAUDE.md.
Always-on rules
Cursor rules set to always apply but written for specific files.
Broken links
Pointers to files that no longer exist, which send your agent looking for nothing.
Unused MCP servers
Servers Claude Code hasn't called in 30 days, so you can switch them off.
Every agent in one audit
One report shows what each agent in your project loads per request, so rules shared between agents are counted once and kept in sync.
| Agent | What it reads |
|---|---|
| Claude Code | CLAUDE.md (project, parent folders, user), imports, .claude/rules, skill descriptions, MCP servers |
| Codex | AGENTS.md, ~/.codex/AGENTS.md |
| Cursor | .cursor/rules (always, described and scoped rules), .cursorrules, AGENTS.md, MCP servers |
| GitHub Copilot | .github/copilot-instructions.md, .github/instructions, AGENTS.md |
| Gemini CLI | GEMINI.md (project and user) |
| Windsurf | .windsurf/rules, .windsurfrules |
| Cline | .clinerules (file or folder) |
See what each Claude Code session costs
The usage check shows this session's tokens, how full the context is, and every time the prompt cache had to be rebuilt, with the likely cause: a model switch, a break longer than the cache lifetime, or a context change. Each rebuild re-sends your whole context at full price.
How it works
Ask
"Put this project on a token diet." Your agent runs the audit and shows the top findings with what each saves.
Approve
Pick the fixes you want. TokenDiet backs up each file, then applies the change exactly as its playbook describes.
Compare
It measures again and shows before and after for each agent. If you use Tokenmeter, the result appears in its dashboard.
On the demo project
The demo project ships with TokenDiet so you can try it safely. Your results depend on your setup; counts are estimates within about ±20%.
Pricing
TokenDiet
$15one-time payment, no subscription
Launching soon on Agensi
- The skill, scripts and playbook
- A demo project to try it safely first
- Updates included: new versions reach you through Agensi
- Use for yourself, or for your whole team or company
- 30-day money-back guarantee
- Support through GitHub, with replies within 2 business days
Sold through Agensi, which handles payment and refunds. A buy button appears here when the listing is live. See the terms and refunds.
Requirements
- An agent that supports Agent Skills (SKILL.md)
- Node.js 18 or newer
- Claude Code, for the session usage check
After you buy
Follow the install guide. It takes about a minute.
Frequently asked questions
Will it make my agent worse?
It is built not to. TokenDiet only removes repetition and noise, or moves instructions so they load when needed. Anything that could affect quality, such as a smaller model or lower effort, is explained and left to you.
Can I undo changes?
Yes. TokenDiet backs up every file before editing it, into a .tokendiet/backup folder in your project, and one command restores them.
Does it send my code anywhere?
No. It runs on your computer with Node.js and reads only your instruction files and Claude Code's local logs.
How accurate are the numbers?
Token counts are estimates within about ±20%, because exact counts depend on each model's tokenizer. They are reliable for finding what is heavy and comparing before and after.
Does it work without Claude Code?
Yes. The audit covers Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf and Cline. The session usage check needs Claude Code, because other agents don't keep comparable local logs.