Pario

Kubernetes-native token cost control plane for LLM APIs.

Pario sits between your applications and LLM providers (OpenAI, Anthropic, etc.) to give you full visibility and control over token spend — budgets, caching, routing, and real-time observability.

Features

• Transparent Proxy — drop-in replacement for OpenAI and Anthropic API endpoints with SSE streaming support
• Token Tracking — per-key, per-model usage tracking with session detection
• Token Budgets — per-team, per-app, per-model spend limits
• Semantic Caching — deduplicate similar prompts (SQLite local, Redis distributed)
• Smart Routing — route requests across models with fallback chains
• Cost Attribution — team/project cost breakdowns with per-model pricing
• Audit Log — opt-in full request/response logging for compliance and debugging
• MCP Server — expose stats, budgets, costs, and audit data to AI agents via Model Context Protocol
• Live Observability — pario top for real-time token usage, Prometheus metrics

Architecture

┌─────────────┐     ┌───────────────────────────────────────┐     ┌──────────────┐
│  Your App   │────▶│                 Pario                  │────▶│  LLM Provider│
│             │◀────│  proxy · budget · cache · router       │◀────│  (OpenAI etc)│
└─────────────┘     └───────────────────────────────────────┘     └──────────────┘
                         │           │           │
                    ┌────▼───┐  ┌────▼───┐  ┌───▼────┐
                    │Metrics │  │ SQLite │  │  CRDs  │
                    │(Prom)  │  │ Cache  │  │(Budget)│
                    └────────┘  └────────┘  └────────┘

Quickstart

# Install (once releases are available)
# brew install pario-ai/tap/pario

# Or build from source
make build
./bin/pario --help

Development

make build   # compile to bin/pario
make test    # run tests
make lint    # run golangci-lint
make clean   # remove build artifacts

License

TBD