I built Governor, a Claude Code plugin for people who hit Max-plan limits during long coding sessions.
Most token-saving tools focus on making the assistant’s replies shorter. That helps, but in real Claude Code sessions the bigger waste often comes from somewhere else:
- huge
CLAUDE.mdand project memory files - noisy test/build logs flooding context
- vague prompts that trigger broad repo scans
- repeated failed fixes
- scope drift during long tasks
Governor tries to solve those problems directly.
What it does
Governor adds:
- compact professional response mode
- safe
CLAUDE.mdcompression - protected-span validation for code, paths, commands, warnings, URLs, and env vars
- Bash/test/log output filtering
- local usage telemetry
- optional planning and drift guardrails
- rule snippets for Codex, Cursor, Gemini, Windsurf, and Cline
The goal is not to make Claude talk in a meme style. The goal is serious context hygiene: keep the agent useful, but stop it from burning quota on avoidable noise.
Why I built it
Long-context coding agents are powerful, but they make it easy to waste context.
A single noisy test command can dump thousands of useless tokens into the session. A bloated memory file gets loaded again and again. A vague prompt like “fix everything” can turn into broad scanning, over-editing, and retry loops.
Governor acts more like a usage governor than a style plugin. It helps measure where quota is going, reduce recurring context, and keep broad tasks from drifting.
Benchmarks
Small local smoke benchmarks:
| Area | Result |
|---|---|
| Output tokens vs control | ~55% saved |
| Memory compression | ~55% saved |
| Synthetic noisy pytest output filtered | ~96% blocked |
These are not universal claims. Governor reports measured local savings instead of fixed magic percentages.
Try it
GitHub:
https://github.com/0xhimanshu/governor
If you use Claude Code heavily and hit usage limits, I’d love feedback from real long-session users.
