# Respan Launches a Rust AI Gateway Built Around Plugins

Respan separates provider protocols from gateway behavior, giving developers more control over the code between an application and its models.

By Paras Madan | 2026-10-21


Respan introduced its open-source Rust AI gateway on October 20, 2026, with a plugin-based design for routing model requests, translating APIs and handling streams. The release targets developers whose provider integrations have grown into their own infrastructure project. It is still pre-alpha, so the architecture matters more than promises of production readiness.

In the [founder's launch post](https://news.ycombinator.com/item?id=50167583), Respan describes a small core with extensible providers, protocols and operational features. The gateway can run without the company's hosted platform, which separately offers managed infrastructure and observability.

## What Respan Gateway supports

The [public repository](https://github.com/respanai/respan-gateway) lists OpenAI Chat Completions and Responses, Anthropic Messages and Gemini generateContent formats. A client can retain its API format while the gateway routes to another provider and translates supported fields.

The scope also includes media endpoints, realtime voice and MCP tools. Routing options include retries, fallbacks and racing requests. The project is dual-licensed under MIT or Apache-2.0.

That feature list should be read as documented support, not proof that every provider-specific behavior survives translation. Teams still need to test the exact combination of tools, streaming events and model options their application uses.

For example, a support assistant might use one provider for routine replies and another for difficult cases. Keeping the client interface stable simplifies that change, but only if both paths preserve the tool-call behavior the application expects.

## How the plugin architecture works

The [plugin guide](https://github.com/respanai/respan-gateway/blob/main/docs/plugins.md) separates wire protocols from typed requests and response events. Plugins can modify, reject or answer requests and filter streamed output. Hooks around upstream calls also cover calls initiated by other plugins.

That separation lets application policy evolve without rewriting protocol adapters. **Plugin order still matters.** A routing decision made before filtering may differ from one made afterward, so modular code still needs integration tests.

## Streaming makes simple filters less simple

The guide warns that a sensitive word can span stream chunks. Replacing text inside each chunk independently can miss it; filtering may need to buffer incomplete matches.

This makes the evaluation workload important. Test interrupted streams, partial tool arguments and client disconnections, not only complete text responses. A gateway is useful only if its behavior remains understandable when requests fail.

## Availability and deployment limits

The [release page](https://github.com/respanai/respan-gateway/releases) lists v0.2.0 on October 20, following v0.1.0 on October 19. That is newer than a README sentence saying no release is tagged. Prefer a pinned release and check its assets rather than following that stale sentence or assuming every package-distribution route is ready.

Without client keys, the documented gateway binds its unauthenticated use to loopback and refuses a public address. That is a starting safeguard, not a complete deployment review. Provider credentials, trace destinations and client authorization still need deliberate configuration.

Respan is worth evaluating for teams that want to own their gateway extensions. The decision should turn on protocol fidelity and operational behavior under failure, not simply on Rust or the number of providers in a README.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn infrastructure launches and developer tooling into clear stories that explain what changes in real workflows.

[Book a 30-minute call](https://cal.com/paras-madan/30min).