contextlab

See what is actually filling your AI coding agent's context window — and what to change to fix it.

Your agent tells you "142,000 tokens, $3.20." It does not tell you that 105,000 of those are one npm install log stuck in the history since turn 3, being re-uploaded on every turn since.

npx contextlab claude       # run your agent through a local proxy
npx contextlab optimize     # then: what to change, ranked by money

A 105,716-token result from Bash has been re-sent 6 times
105,716 tokens × 6 turns = 528,580 wasted  ·  $2.67

FIX
Start a fresh session once you have taken what you need from this output.

Commands

| Command | Does | | :--- | :--- | | contextlab <tool> | run a coding agent through the proxy | | contextlab optimize | ranked waste + the exact config change | | contextlab why | why was my last turn expensive? | | contextlab cost | historical spend, by day or project | | contextlab watch | live context gauge in the terminal | | contextlab dashboard | the web UI on localhost:4041 | | contextlab doctor | check everything is ready |

Agents

Claude Code, Aider, Gemini CLI and Copilot CLI work with nothing installed. Codex, Cline and OpenCode cannot be redirected by an environment variable and need brew install mitmproxy — contextlab doctor checks the certificate too.

Nothing leaves your machine

No account, no cloud, no telemetry, and no LLM is called anywhere in the product. Every number is arithmetic; every recommendation is a deterministic rule. It works offline.

The proxy is the only component that sees a credential, so it is ~560 lines with zero dependencies — short enough to read before trusting it. Header values are destroyed before any capture reaches disk.

Full documentation, architecture and the export format: https://github.com/eyekaushal/contextlab

MIT