654 stars in 18 days: Antirion, created on October 17, 2026, is one of the fastest-rising new repositories on GitHub. It offers what many platform teams are building by hand: a single point between your applications and every model provider, with routing, budgets, guardrails and metrics. What stands out is the packaging, one Rust binary with the console built in, and the number that comes with it: 0.53 ms median overhead at 5,000 requests per second on 4 vCPUs, according to its author's benchmark.
What it is
An AI gateway under the Apache-2.0 license. Applications talk to it with the OpenAI, Anthropic, Gemini or Bedrock SDKs, and Antirion translates between dialects, so an OpenAI Chat client can end up on a Claude deployment and vice versa, streaming included. Behind it sit OpenAI, Azure OpenAI, Anthropic, Bedrock, Vertex AI, Mistral, Groq, DeepSeek and any compatible server such as vLLM.
Every request goes through a fixed pipeline: authentication with virtual keys or JWTs, rate limits and budgets, input guardrails, exact or semantic caching, routing, attempts with retries and circuit breakers, and output guardrails with an entry in a cost ledger. Routing supports fallback, weighted splits, least cost, least latency or least busy, and a simulator tells you which deployment a request would reach and why. Budgets are set in dollars or tokens per organization, team, project or key, and can be soft, hard or held for approval.
For operations it ships what you expect from something on the critical path: Prometheus metrics, OTLP traces and logs, alerts with SLOs, a distroless non-root image on a read-only filesystem, and binaries with an SBOM, Sigstore signatures and SLSA provenance. It runs as a single node with an embedded store or as a cluster on PostgreSQL and Valkey, with a Helm chart. It also acts as an MCP gateway and an A2A proxy for agents.
Trying it
The README starts with Docker and an OpenAI or Anthropic key, no config file needed:
docker run -p 8080:8080 -p 8081:8081 \
-v antirion-data:/var/lib/antirion eliasbiondo/antirionThe first start prints the console address and an admin token that is shown only once (it is kept in the volume). From the console you add the provider, a model with your own name for it (say chat) and a virtual key, and your apps point their base URL at http://localhost:8080/v1:
curl -s http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer $ANTIRION_API_KEY" -H 'Content-Type: application/json' \
-d '{"model": "chat", "messages": [{"role": "user", "content": "Hello"}]}'Port 8080 is the data plane and 8081 the admin plane (API and console). Anywhere other than your laptop, 8081 should not be exposed the way 8080 is.
What I like and what I don't
I like the design. Routing rules, budgets, guardrails and a request log in one binary, with configuration as a versioned YAML document with diffs and rollback, is exactly what you want from a platform component. And the supply chain is taken seriously from day one, which matters in a component that holds the keys to all your providers.
What I don't: maturity. The latest published image is 0.1.1, from October 18, 2026; the project has 20 open issues and, as far as the repository shows, a single author. The feature list reads like a product several years old, and at two weeks old that calls for careful testing before it sits in front of production traffic. The benchmark is the author's own, although the method and the harness to rerun it on your AWS account are published.
Antirion vs LiteLLM
LiteLLM is the open source reference here, and Antirion's benchmark uses it for comparison: with LiteLLM 1.103.1 on the same machine, every request failed at 5,000 requests per second, and its ceiling was under 500. Those figures are for that specific release and a minimal setup (one key, no logging, no cache), not a typical deployment.
| Antirion | LiteLLM | |
|---|---|---|
| Approach | Single Rust binary with console | Proxy (AI Gateway) and Python SDK |
| License | Apache-2.0 | MIT, except the enterprise/ directory, under a commercial license |
| Deployment | Distroless Docker, binary, Helm; single node or cluster | Self-hosted proxy; enterprise edition with SOC 2 Type II |
| Maturity | Created October 2026, version 0.1.1, 654 stars | In development since 2023 |
When to use which: LiteLLM if you need a broad provider catalog, commercial support and a tool with three years behind it. Antirion if gateway latency and footprint matter to you, you want everything in one binary, and you can afford to adopt something very young, starting in an environment without critical traffic.
Sources
- eliasbiondo/antirion (GitHub)
- eliasbiondo/antirion image tags (Docker Hub)
- LiteLLM license (BerriAI/litellm)
- LiteLLM Enterprise (LiteLLM Docs)