AIAI NO. ---

OpenAI donates Open Responses API spec to Agentic AI Foundation

The spec for a common LLM API now sits beside MCP and AGENTS.md under neutral governance. Ollama and LM Studio already expose compatible endpoints, but only 26 of 120 surveyed providers declare it.

By Rowan Ashby Published 23.10.26 13:16 UTC 6 min read

The Linux Foundation / Wikimedia Commons / Public domain · Source

Key takeaways

OpenAI has donated Open Responses, its open specification for a common LLM API, to the Agentic AI Foundation (AAIF), a Linux Foundation project. The move was announced on Oct. 22, 2026, at AGNTCon + MCPCon North America in San Jose.
Three other projects joined in the same announcement: Anthropic's Agent Skills, Hugging Face's OpenEnv and Alibaba's OpenSandbox. AAIF now hosts 10 projects, up from three at its launch in December 2025.
The Open Responses API is based on OpenAI's Responses API. It ships a written specification, an OpenAPI reference and an acceptance-test suite that any provider can run against its own endpoint.
Local runtimes already expose it. Ollama documents a stateless /v1/responses endpoint from v0.13.3, and LM Studio made its endpoint Open Responses-compatible in version 0.3.39.
Adoption is still thin. A September survey by API Evangelist found that 26 of 120 OpenAI-compatible providers declare POST /responses in their published OpenAPI files.

What the Open Responses API is and who now owns it

OpenAI has donated the Open Responses API, its open specification for a common LLM interface, to the Agentic AI Foundation (AAIF), a Linux Foundation project. The move was announced on Oct. 22 at AGNTCon + MCPCon North America in San Jose.

OpenAI first published Open Responses on Jan. 15, 2026, in an announcement on its developer community forum. Until this week the spec was led by one vendor. The project's website now carries the notice that it is "a Series of LF Projects, LLC", the Linux Foundation's legal structure for hosted projects, and it links to a technical charter that sets out how decisions are made.

Steve Coffey, a member of OpenAI's technical staff, said in the release that the project aims to give builders "simple, clean, and unified interfaces" for language-model products.

What the specification contains

The spec, currently dated version 2026-04-24 (the previous version was 2026-01-15), defines a small set of shared building blocks:

Items. The basic unit of context. An item can be a message, a function call, or a reasoning record. Every item has an id, a type and a status (in_progress, incomplete or completed).
Streaming events. Server-sent events come in two kinds. Delta events such as response.output_text.delta carry new content. State events such as response.completed or response.failed mark changes in status. Every stream must end with a terminal response event.
Tools. The spec separates externally hosted tools, such as function calls and MCP servers that the developer runs, from tools hosted inside the provider, such as file search. tool_choice and allowed_tools control which tools the model may call.
State. previous_response_id lets a server continue a conversation from a stored response. Version 2026-04-24 added a WebSocket transport that allows one in-flight response per connection, with connections capped at 60 minutes.
Extensions. Providers can add their own item types, events and hosted tools, but these must carry a provider prefix (for example openai:web_search_call) and must not change core behaviour.

The acceptance tests are the most practical part. On the project's compliance page, a developer enters a base URL, a model name and an API key, and the suite checks the endpoint against the OpenAPI schema. It runs 10 tests in the browser, covering basic text, streaming, system prompts, tool calling, image input, multi-turn chat and the compaction endpoint. Seven more tests, run from the command line, cover the WebSocket transport. The page does not yet publish a list of implementations that pass.

Which runtimes support Open Responses

The project's website shows a "Community" section with the logos of Nvidia, Vercel (AI SDK), OpenRouter, Hugging Face, LM Studio, Databricks, Red Hat, AWS, Ollama, OpenAI, vLLM and Llama Stack. The site does not say what role each organisation plays. The table below lists only what Bullquill could confirm from each project's own documentation.

Runtime	Listed on openresponses.org	What its own documentation says
Ollama	Yes	`/v1/responses` added in v0.13.3; stateless only (no `previous_response_id` or `conversation`); supports streaming, function tools and reasoning summaries
LM Studio	Yes	Partnered with OpenAI on Open Responses; version 0.3.39 (Jan. 15, 2026) made its `/v1/responses` endpoint compatible with the spec
vLLM, OpenRouter, Hugging Face, Vercel, Nvidia, AWS, Databricks, Red Hat, Llama Stack	Yes	Shown as community logos; the level of support was not checked for this article

Sources: openresponses.org, Ollama OpenAI-compatibility docs and the LM Studio blog, checked Oct. 23, 2026.

In practice, a single request shape now works against a local server and a hosted one. Here is a minimal call to a local Ollama instance (see our guide to running a local LLM with Ollama):

curl http://localhost:11434/v1/responses \
  -H "Content-Type: application/json" \
  -d '{"model": "llama3.2", "input": "Explain the Open Responses API in one sentence."}'

Pointing the same body at another compliant base URL should, in principle, return the same item structure. That is the promise the acceptance tests are meant to check.

How Open Responses fits next to MCP and AGENTS.md

AAIF already hosts the Model Context Protocol (MCP), which standardises how agents connect to tools and data, as well as AGENTS.md, the goose agent and the Agent2Agent (A2A) protocol. Open Responses covers a different layer: the call from an application to the model itself. Agent Skills, which Anthropic first released in December 2025, packages instructions, scripts and resources that an agent loads when needed.

AAIF project	Contributed by	Layer it standardises
Open Responses	OpenAI	Application-to-model API calls, streaming and tool-call format
Model Context Protocol (MCP)	Anthropic	Connecting agents to external tools and data
Agent Skills	Anthropic	Packaging reusable agent capabilities
OpenEnv	Hugging Face	Reinforcement-learning environments for agents
OpenSandbox	Alibaba	Sandboxed runtimes for running agent code

Source: Linux Foundation press release, Oct. 22, 2026.

Manik Surtani, AAIF's CTO and co-founder, said the foundation is "giving the community one place to build shared standards, protocols, and infrastructure." Developers building MCP servers in Python or tool-calling agents now have both the tool layer and the model layer under one governance body.

How far adoption has got

API Evangelist, which tracks published API definitions, measured provider interfaces on Sept. 13, 2026. Its figures show how much ground the spec still has to cover against the older Chat Completions format.

Operation (OpenAI-compatible providers with readable OpenAPI, n = 120)	Providers declaring it	Share
`POST /chat/completions`	105	87.5%
`GET /models`	86	71.7%
`POST /embeddings`	39	32.5%
`POST /responses`	26	21.7%
`GET /responses/{response_id}`	9	7.5%

Source: API Evangelist, data measured Sept. 13, 2026, published Oct. 23, 2026.

The same survey counted 34 providers that claim compatibility with Anthropic's Messages API, of which 29 publish an OpenAPI file. The author argued that the spec succeeds only if it also maps cleanly onto those Messages-style providers. Otherwise, the author said, the market stays split into two camps with adapters in between.

What it means

For developers, little changes overnight. Code written against OpenAI's Responses API should keep working, and the spec was already public. What changes is ownership: decisions about the format now go through a foundation charter, not a single company's roadmap, which lowers the risk for rival providers that adopt it.

The practical benefit is less glue code. Teams that today maintain separate clients for OpenAI, Anthropic and local models could target one schema for messages, tool calls and streaming, then switch providers by changing a base URL. That also makes it easier to route traffic to cheaper models, a technique covered in our guide on how to reduce LLM API costs.

Three open questions remain. Feature support varies: Ollama, for example, does not support stored conversation state. No public pass list from the acceptance tests exists yet. And Anthropic, which contributed MCP and Agent Skills to the same foundation, has not said its own API will adopt Open Responses. Until providers publish test results, developers should run the compliance suite against any endpoint they plan to rely on.

Frequently asked questions

What is the Open Responses API?

Open Responses is an open specification for a common LLM API, based on OpenAI's Responses API. It defines shared items, streaming events and tool-call formats so one request can run across different model providers, and it ships an OpenAPI reference and acceptance tests.

Who owns Open Responses now?

OpenAI contributed it to the Agentic AI Foundation, a Linux Foundation project, on Oct. 22, 2026. The project is run under a technical charter as a Series of LF Projects, LLC.

Does Ollama support the Open Responses API?

Ollama documents a /v1/responses endpoint from version 0.13.3. It is stateless only, so previous_response_id and conversation are not supported, but streaming, function tools and reasoning summaries are.

How is Open Responses different from MCP?

MCP standardises how agents connect to external tools and data. Open Responses standardises the call from an application to the model itself. Both are now hosted by the Agentic AI Foundation.