token-diet
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@token-dietInitialize the repo map and enable diff-only mode for this project."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
token-diet 🍽️
Save tokens while coding with AI. Use less, get the same (or better) results.
When you use AI coding tools like Claude Code, Codex, or Cursor — they send your entire codebase and chat history to the AI every single time. That's expensive and actually makes the AI dumber (too much noise buries the important stuff).
token-diet fixes that. It sends a smart summary of your code instead of every file, remembers decisions instead of replaying the whole conversation, reuses cached prompts so repeated parts are nearly free, and makes the AI reply with just the changed lines instead of rewriting entire files.
The result: you spend less on tokens, and the AI gives better answers because it can focus on what matters.
Results
What we measured | Result |
🔽 Input tokens saved | ~70% fewer tokens sent to the AI |
🔽 Output tokens saved | ~65% fewer tokens in AI replies |
✅ Code quality kept | ~97% of tasks still pass correctly |
⚡ Extra time added | ~60 ms per request (barely noticeable) |
Related MCP server: parecode
Architecture
flowchart TB
subgraph Client["Your AI tool"]
CC["Claude Code / Cowork / Codex / Cursor"]
end
subgraph Frontends["How you connect"]
MCP["Slash commands (/map, /focus, etc.)<br/>Works in Claude Code, Cowork"]
PROXY["Local proxy (localhost:8000)<br/>Works with any AI tool"]
end
subgraph Engine["What happens behind the scenes"]
ASM["Prompt builder<br/>Stacks info in the smartest order"]
BUD["Token counter<br/>Measures exact cost, cuts the fat"]
IDX["Code scanner<br/>Reads your files, extracts structure"]
MAP["Smart code map<br/>Ranks what's important right now"]
MEM["Memory<br/>Remembers decisions, forgets chit-chat"]
DOC["Doc converter<br/>Turns PDFs/Word docs into clean text"]
PATCH["Edit applier<br/>Takes AI's changes, patches your files"]
ROUTE["Model picker<br/>Uses cheap AI for simple tasks"]
TEL["Cost tracker<br/>Shows how much you saved"]
end
subgraph Store["Saved data (per project)"]
DB[("Code index<br/>SQLite database")]
LOG[("Decision log<br/>What was decided & why")]
CACHE[("Cache<br/>Avoids re-doing work")]
end
PROV["AI provider<br/>Anthropic / OpenAI"]
CC -->|"your message"| MCP
CC -->|"your message"| PROXY
MCP --> ASM
PROXY --> ASM
ASM --> BUD
ASM --> MAP
ASM --> MEM
ASM --> DOC
MAP --> IDX
IDX --- DB
MEM --- LOG
DOC --- CACHE
ASM -->|"optimized prompt"| ROUTE
ROUTE --> PROV
PROV -->|"just the changes"| PATCH
PATCH -->|"edits applied"| CC
ROUTE --> TEL
TEL --- CACHEHow it flows: Your message → token-diet builds a lean prompt (code map + decisions + just what's needed) → picks the right AI model → AI replies with only the changed lines → edits applied to your files. Every request is tracked so you can see your savings.
Install
Quickest way (from GitHub):
pip install 'git+https://github.com/aryxnsdfs/token-diet'With all features:
pip install 'token-diet[all]'For development:
git clone https://github.com/aryxnsdfs/token-diet
cd token-diet
pip install -e '.[all,dev]'Quick start
cd your-project # go to any project you're working on
ctx init # one-time setup: scans your code, creates config
ctx doctor # check everything is workingThen open Claude Code in that project and press / — you'll see the
commands. Type /begin as your first message to begin a session.
Commands
Command | What it does |
| Begin a coding session — scans your code, shows the overview, turns on smart mode |
| Show the whole project — a ranked overview of your code (not every file) |
| Open a file — pulls the full file into the conversation |
| Find one function or class — shows just that piece of code |
| Keep replies short — AI only writes the lines it's changing |
| Free up chat space — summarizes old messages into key decisions |
| Show savings — how many tokens saved, cache hits, money saved |
| Switch AI models — use a cheap/free local model for simple tasks |
| Apply edits — applies the AI's code changes to your files |
Works with
Tool | How to connect |
Claude Code (terminal or VS Code) | Just run |
Claude Cowork / Claude.ai | Run |
Codex / Cursor / any tool | Run |
See docs/INSTALL.md for detailed setup per tool.
How it saves tokens
Technique | What it does | Tokens saved |
Code map | Shows structure (function names, classes) instead of full files | ~86% of input |
On-demand pull | Only loads a file when you ask for it with | Avoids unnecessary files |
Decision log | Remembers "we chose bcrypt for auth" instead of replaying 50 messages | ~80% of old chat |
Prompt caching | Reuses the unchanged part of the prompt (nearly free to resend) | ~10x cheaper reads |
Diff-only output | AI writes 5 changed lines, not the whole 500-line file | ~95% of output |
Model routing | Uses a small cheap model for simple tasks (summaries, commit msgs) | Uses free local AI |
Project structure
ctx/
cli.py Command line: init, serve, doctor, proxy
registry.py All commands defined in one place
server.py Connects to Claude Code (MCP protocol)
proxy.py Connects to any tool (local web server)
init/ Setup scripts for each AI tool
engine/ The brain: code scanner, map builder, memory, etc.
tests/ Automated tests to make sure nothing breaksLicense
MIT — use it however you want.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Latest Blog Posts
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/aryxnsdfs/token-diet'
If you have feedback or need assistance with the MCP directory API, please join our Discord server