Agentic API is an open-source, stateful API layer for agentic applications built on vLLM, written in Rust by the vLLM project. Published on GitHub under the Apache-2.0 license with about 330 stars and updates in October 2026, it runs OpenAI-grade agentic workloads on your own GPUs.
vLLM delivers high-throughput inference, but real agentic applications need more than raw tokens: they need conversation state, tool-call loops, and multi-turn orchestration, which today lives in client code. Agentic API moves that complexity server-side. It is a Rust-native gateway that sits in front of vLLM and owns the stateful agentic APIs, starting with an OpenAI-compatible Responses API and a Messages API. Your application makes one API call and the server handles state hydration through previous response identifiers, server-side tool execution, streaming over HTTP, SSE, and WebSocket, and continuation. Persistence is handled with SQLite or PostgreSQL, and supported clients include coding agents such as Codex and Claude Code.
The project is explicit that vLLM is one supported backend rather than part of the product itself: the gateway owns the agentic semantics, while inference stays on the serving stack a team already runs. That separation is what lets the same client code, whether it comes from a coding agent, an SDK, or plain curl, target a self-hosted deployment through familiar Responses and Messages API shapes.
Agentic API is free and open source under the Apache-2.0 license. There is no hosted tier; you deploy the gateway yourself in front of your own vLLM inference servers.
It is for platform and infrastructure teams self-hosting open models with vLLM who want managed-agent API semantics, state, tools, and streaming, without sending workloads to a proprietary API.