Route every AI call to the cheapest model that can do the job well.
48 MCP tools · 20+ LLM providers · intelligent routing · personal memory · budget tracking · decision analytics.
Result: 60–80% cost reduction vs running everything on Claude Opus.
Table of Contents
The Problem & Solution
The Problem
Traditional AI-assisted development routes every task to your most capable (and expensive) model. A simple file lookup costs the same as complex architecture redesign—burning through your quota and budget on low-value work.
The Solution
llm-router analyzes each task and routes it to the cheapest model that can handle it well:
- Simple lookups → Ollama/Haiku (free/cheap)
- Moderate coding → Gemini Pro/GPT-4o (budget-friendly)
- Complex reasoning → Claude Opus/o3 (premium, only when needed)
The magic: You keep the same conversational experience. No manual routing, no model picking. It just works behind the scenes.
Real-World Savings
Proven Results from 14-Day Sprint
Real numbers: 51 releases, 22.6M tokens, $6.95 spent (vs $50–60 with traditional Opus-everywhere approach).
Cost Breakdown
- Actual spend: $6.95 (22.6M development tokens)
- Opus baseline: $50–60 (traditional approach)
- Savings: $43–53 per 2 weeks (87% reduction)
- Annualized: ~$180/year vs $1,200–1,500 baseline
Token Distribution (Free-First Routing)
31% from free models → 7.0M tokens, $0 cost (Ollama + Codex)
38% from budget models → 8.6M tokens, $2.82 cost (Gemini Flash + GPT-4o-mini)
31% from premium models → 7.0M tokens, $4.13 cost (GPT-4o, Pro, Claude)
Quota Pressure Elimination
Free-first routing eliminated budget pressure over the sprint—enabling sustainable feature velocity without cost anxiety.
Quick Start
1. Install
# One command to install and configure
pipx install llm-routing && llm-router install
2. (Optional) Add Provider Keys
# Available in .env or ~/.llm-router/config.yaml
export OPENAI_API_KEY="sk-..." # For GPT-4o, o3 (optional)
export GEMINI_API_KEY="AIza..." # For Gemini models (free tier available)
export OLLAMA_BASE_URL="..." # For local Ollama (optional, auto-starts)
3. Done
Start using Claude Code, Gemini CLI, Codex, VS Code, Cursor, or any MCP-compatible editor. llm-router handles everything automatically.
Key Features
🎯 Intelligent Routing
- Complexity Classification — Analyzes prompts to determine if task is simple/moderate/complex
- Provider Fallback Chains — Ollama → Codex → GPT-4o → Claude (free-first, always)
- Budget Pressure Awareness — Automatically downgrades model when quota is limited
- Quality Monitoring — Demotes models with degraded performance via judge scoring
- Decision Logging — Track which model was selected, why, and the cost impact