Multi-provider, capability-aware, cost-optimized model routing for AI agents.

Routes tasks to the optimal LLM based on complexity classification, required capabilities, cost constraints, and latency requirements.

• Multi-provider: Anthropic, Google, OpenAI, Open-source (via OpenRouter)
• 12 models in registry with full cost/capability metadata
• Keyword-based complexity classifier (0-20 score)
• Capability filtering: require specific skills (reasoning, code_review, etc.)
• Budget-aware ranking: low/medium/high cost sensitivity
• Overkill prevention: penalizes expensive models for simple tasks
• Fallback chain: cross-provider failover
• Decision logging: JSONL output for learning from outcomes
• CLI + Library: use standalone or import into agents

# Simple routing
python3 scripts/router.py --task "Check disk space" --budget low
# → gemini-2.5-flash ($0.15/MTok)

# Complex task
python3 scripts/router.py --task "Design microservice architecture" --budget high
# → claude-opus-4-6 ($5.00/MTok)

# Require capabilities
python3 scripts/router.py --task "Review PR" --require code_review --provider anthropic
# → claude-sonnet-4-6

# JSON output for scripting
python3 scripts/router.py --task "Debug race condition" --json

# Verbose with alternatives
python3 scripts/router.py --task "Analyze logs" -v

Task → Complexity Classifier (keywords + length) → Score 0-20
     → Capability Filter (hard requirement)
     → Budget-Weighted Ranking (quality × cost × speed)
     → Overkill Penalty (cheap tasks don't get expensive models)
     → Pick best + 2 alternatives

• Light (0-2): status checks, formatting, simple Q&A
• Medium (3-8): code review, debugging, analysis, implementation
• Heavy (9+): architecture, security audit, novel problems

python3 scripts/benchmark.py

Current: 15/15 (100%) across light/medium/heavy task categories.

import sys
sys.path.insert(0, "scripts")
from router import pick_model, load_models, classify_complexity

models = load_models("config/models.json")

result = pick_model(models, "Design auth system", budget="high")
print(result["model"]["name"])       # claude-opus-4-6
print(result["model"]["hermes_id"])  # anthropic/claude-opus-4-6
print(result["tier"])                # heavy
print(result["score"])               # 15

Edit config/models.json to:

• Add new models
• Update pricing
• Adjust power scores
• Add/remove capabilities

Works with delegate_task — route before spawning sub-agents:

result = pick_model(models, task_description, budget="medium")
# Use result["model"]["hermes_id"] as the model parameter

MIT