What Is Open Responses? The AAIF Multi-Provider LLM API Spec
The short answer
Open Responses is an open specification for one LLM API that works across model providers, based on OpenAI’s Responses API. It defines a shared schema for requests and outputs, consistent streaming events and tool calling, so an application can describe a request once and run it against different providers. On October 22, 2026, OpenAI contributed it to the Agentic AI Foundation (AAIF), the Linux Foundation body that already hosts MCP, putting it under vendor-neutral governance.
What the spec defines
According to the project’s site and repository, Open Responses centres on four ideas:
- Items as the atomic unit of context and model output — messages, tool calls and tool results — each with a clear state machine.
- An agentic loop: the model emits tool calls, receives results and continues, inside one interface.
- Semantic streaming events rather than raw text deltas, so clients behave the same regardless of provider.
- Extensibility without fragmentation: a stable core with room for provider-specific tools and item types that do not generalise yet.
The spec is written in TypeSpec, published as OpenAPI, and released in immutable dated versions (the site lists 2026-01-15 and 2026-04-24 releases). A compliance test suite — interactive at openresponses.org/compliance and as a CLI — checks an endpoint against it.
Why it matters for builders
Most teams already abstract providers with an OpenAI-compatible Chat Completions layer, a gateway or an SDK. That works for plain text but leaks as soon as you use tools, reasoning items, multimodal input or streaming, because every provider encodes those differently. A written spec with acceptance tests gives three practical benefits:
- Portability you can verify. A provider or gateway can claim compliance and you can test the claim.
- Cleaner agent code. Tool calls and their results are first-class items, not strings to parse.
- Less lock-in to one vendor’s SDK. Neutral governance means OpenAI no longer controls the spec alone.
Support is still early. Open WebUI, for example, documents Open Responses support as experimental. Check your provider or gateway — OpenRouter, LiteLLM and similar routers — before you rewrite anything.
The other three AAIF projects announced the same day
| Project | Contributed by | What it is |
|---|---|---|
| Agent Skills | Anthropic | Open format for packaging instructions, scripts and resources that agents load on demand |
| Open Responses | OpenAI | Multi-provider LLM interface spec |
| OpenEnv | Hugging Face | Framework and hub for interoperable reinforcement-learning environments, exposing MCP and the Gym training protocol |
| OpenSandbox | Alibaba | Sandbox platform for agents with multi-language SDKs and Docker and Kubernetes runtimes |
With these, AAIF hosts 10 projects in its first year, up from three at launch in December 2025; the others are MCP, Agent2Agent (A2A), AGENTS.md, goose, agentgateway and Agent Router.
Should you adopt it now?
- Building a gateway, SDK or model-serving API: yes — implement it and run the compliance suite; it is becoming the shared target.
- Building an app on one provider: no rush. Use your provider’s Responses-style API and keep an eye on compliance claims.
- Building agents across several providers: prototype against an Open Responses–compliant layer so tool calls and streaming behave the same everywhere.
Related: agent plugins vs MCP vs agent skills and how to install Claude Code plugins and skills.
Last verified: October 26, 2026.