I built a tool that cuts Claude/ChatGPT token usage by 97% — here's how it works

The Problem

You're debugging a bug. You open Claude. You paste 10 files. You hit the context limit.

Or worse — you paste the wrong files and Claude gives you a useless answer.

This happens because most repos have hundreds of files but you only need 10-15 for any given task.

What I Built

ContextOS — it scans your entire repo, ranks files by relevance to your current task, and exports a perfectly-sized context pack. Secrets are automatically redacted. No cloud. No accounts.

Real Numbers

Tested on FastAPI (2,811 files):

bash
git clone https://github.com/tiangolo/fastapi
cd fastapi
pip install rm-contextos
contextos scan
contextos pack --task "add rate limiting to auth endpoint"

Output:
Without ContextOS : ~284,000 tokens  (2,811 files)
With ContextOS    : ~7,998 tokens    (15 files)
Saved             : ~276,002 tokens  (97% reduction)

How It Works

1. contextos scan — indexes every file: imports, exports, symbols, purpose
2. contextos pack --task "your task" — ranks files using:
  - Keyword matching against your task description
  - Import graph centrality (files imported by many others rank higher)
  - AST symbol extraction (matches function/class names)
  - Git churn scoring (recently modified files rank higher)
3. Fills the token budget greedily, redacts 14 secret patterns, exports

MCP Server

It also runs as an MCP server so Claude Desktop/Code can call it directly:

pip install "rm-contextos[mcp]"
contextos serve --stdio /path/to/project

Claude then has tools: pack_context, list_files, get_file, churn_report.

Install

pip install rm-contextos        # base
pip install "rm-contextos[all]" # + MCP + AST + headroom compression

GitHub: https://github.com/Rohithmatham12/ContextOS
Docs: https://Rohithmatham12.github.io/ContextOS/

Apache-2.0. Fully local. PRs welcome.