Plano is an AI-native proxy server and data plane for agentic applications, built by Katanemo on top of Envoy. It moves the messy parts of LLM integration — smart model routing, guardrails, observability and orchestration — out of your agent code and into dedicated infrastructure. The project has about 7,100 stars on GitHub and is actively developed in 2026.
Plano (the successor to Katanemo’s Arch project) sits between your application and model providers. Incoming prompts pass through guardrails that can block jailbreaks and harmful content; an intent-aware routing layer sends each request to the right model — cheap and fast for simple tasks, frontier models for hard ones — and can call tools and functions as part of the flow. Because it is built on Envoy, Plano inherits production-grade networking: retries, timeouts, load balancing and OpenTelemetry-based tracing and metrics for every LLM call, so you can see cost, latency and errors per route. It speaks the OpenAI API format, supports providers such as OpenAI, Anthropic, Gemini, Mistral and locally hosted models via Ollama or vLLM, and is configured declaratively rather than in application code. The result is that agent developers keep their core logic clean while platform concerns live in one governed layer.
Plano is free and open source under the Apache-2.0 license for self-hosting. You pay only for the model APIs you route through it.
Plano is for platform and agent developers who want routing, safety and observability for LLM traffic handled once, in infrastructure, instead of re-implemented inside every agent.