optym-code
Stop burning expensive AI quota on cheap tasks. Smart model routing for Claude Code and Codex CLI.
The Problem
Claude Code and Codex charge per request — or burn through quota fast. Most tasks don't need your most expensive model.
Without optym-code:
"hello" → Opus / o3 (expensive, wasted)
"read this file" → Opus / o3 (expensive, wasted)
"add try-catch" → Opus / o3 (expensive, wasted)
"design architecture" → Opus / o3 (actually needed)
Result: quota gone in 2 hours
With optym-code:
"hello" → Haiku / gpt-4.1-mini (20x cheaper)
"read this file" → Haiku / gpt-4.1-mini (20x cheaper)
"add try-catch" → Sonnet / gpt-4.1 (5x cheaper)
"design architecture" → Opus / o3 (escalated)
Result: same work, 30-55% less cost
Works as a local proxy — invisible to your CLI. No config beyond install, no data leaves your machine.
Install
curl -s https://raw.githubusercontent.com/arturoyo/optym-code/master/install.sh | bash
Restart Claude Code. Done.
Codex CLI Support
optym-code setup codex
optym-code start
Writes ~/.codex/config.toml pointing Codex at the proxy. Requires OPENAI_API_KEY set in your environment.
How It Works
A local proxy runs on localhost:8088. Your CLI sends requests through it.
- Classifies prompt complexity (local regex rules — instant, free, no data sent)
- Routes to cheapest model that can handle the task:
- Haiku / gpt-4.1-mini — greetings, file reads, git ops, simple questions
- Sonnet / gpt-4.1 — moderate tasks, default for ambiguous prompts
- Opus / o3 — architecture, complex debug, multi-file refactors, code with context
- Forwards directly to Anthropic or OpenAI (your keys, your auth)
- Tracks savings in local SQLite
Pro mode upgrades classification to ML (92% accuracy, 45-55% savings) via api.optym.pro. Your LLM traffic never goes through optym servers — only the classification request does.
Compatibility
| CLI | Supported | Setup |
|---|
| Claude Code (subscription) | Yes | install.sh |
| Claude Code (API key) | Yes | install.sh + set ANTHROPIC_BASE_URL |
| Codex CLI | Yes | optym-code setup codex |
| Aider, Cursor, any Anthropic client | Yes | Set ANTHROPIC_BASE_URL=http://localhost:8088 |
Savings
Real numbers from production usage:
| Mode | Routing accuracy | Typical savings |
|---|
| Free (local rules) | ~70% | 20-35% |
| Pro (ML cloud) | 92% | 45-55% |
Commands (Claude Code)
| Command | What |
|---|
/optym | Activate terse mode (cuts output tokens ~65%) |
/savings | Show savings dashboard |
/force-opus | Force Opus for all requests |
/force-haiku | Force Haiku for all requests |
/force-sonnet | Force Sonnet for all requests |
/optym-reset | Return to auto-routing |
stop optym | Deactivate terse mode |
Manual proxy control
optym-code start # start proxy (auto-started on Claude Code session)
optym-code stop # stop proxy
optym-code status # show proxy status and savings
optym-code setup codex # configure Codex CLI
Optym Pro
Free tier: static regex rules, 100% local, 20-35% savings.
Optym Pro — ML classification, 92% accuracy, 45-55% savings. $9/month.
export OPTYM_PRO_KEY=optym_your_key
Your LLM traffic stays direct to Anthropic/OpenAI. Only prompt text is sent to classify.
Privacy
- Free mode: 100% local. Nothing leaves your machine.
- Pro mode: prompt text sent to
api.optym.pro/v1/classify for ML routing. Not stored.
- Telemetry: anonymous daily ping — install count, model distribution, savings %. No prompts.
- Opt-out:
{"telemetry": false} in ~/.optym-lite/config.json
License
MIT
JuliusBrussee
thedotmack
xrf9268-hue
alirezarezvani