[Enhancement]: Support for OpenAI's new Responses API

[PierreMesure]

What features would you like to see added?

Problem Description

OpenAI announced last week a new Responses API that contains the feature set of the Chat Completions API and replaces the Assistants API (the latter is now deprecated and will be sunseted mid-2026).

Azure already announced that they provide this API day one and considering OpenAI's position on the market, we can expect that it will become an industry standard as is the Chat Completions API today.

Desired Solution you'd like

• Ability to connect LibreChat to AI products exposed through the Responses API to interact with them in LibreChat
  - Ability to create and configure OpenAI assistants and agents in the LibreChat UI in the same way that it is today possible to configure assistants through the Assistants API in the UI. (Removed because irrelevant)

More details

• Official documentation for the Responses API
• Thread on OpenAI's developer forum
• Comparison between the Chat Completions API and Responses API
• A blog post from Simon Willison on the new API
• A thread on X by one of the creators of the API talking about its making

Which components are impacted by your request?

Endpoints

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

We have our own agent framework: https://www.librechat.ai/docs/features/agents

While our foray into Assistants API was fruitful, it became non-trivial to maintain and never allowed the flexibility we wanted to see from an agentic framework. I also suspected it would be deprecated and so motivation to maintain was low.

We don't want to be vendor-locked and want high-precision flexibility for agents. All LLM interactions within LC will migrate to the underlying agents framework. API access to LC agents is also planned.

In light of all this, working on integrating Responses API seems like a repeat of Assistants API, and would take away from developer time on agentic experiences for all major providers, in favor of an OpenAI product.

[PierreMesure]

Hi @danny-avila, thank you for your answer! I completely understand your choice of not providing support for proprietary agents standards. I removed this point in my description above.

That being said, LibreChat uses OpenAI's Chat Completions API spec as its standard to interact with custom endpoint providers. If the new Responses API becomes a new standard, it would be interesting to support it for custom endpoints.

A benefit I see is that we build RAG-assistants with LlamaIndex and integrate them in LibreChat using LitServe and the OpenAI API spec (Chat Completions). But we cannot pass RAG references this way as it isn't part of the spec. Responses seems to provide this possibility.

But it's very early to tell whether it will be adopted by 3rd parties and Chat Completions API will most likely remain a standard for many years to come. So I don't mind you closing this issue, it will remain a documentation of the state of things regarding the Responses API for anyone interested. And it can always be reopened when it becomes relevant. 🙂

[danny-avila]

@PierreMesure just want to say, I would be open to use responses via our agents framework somehow, similar to Bedrock, which relies on the agent framework as its backbone, but not to the same extent that we did with assistants api.

[cranberry3148]

The new o1-pro model only supports the Responses API

[PierreMesure]

I didn't know that!

Here's a source and a nuance, they are considering adding it to Chat Completions API.

[danny-avila]

I'm re-opening this, because we can integrate this by simply benefiting from the tools OpenAI offers for this, rather than integrating this wholesale as a separate endpoint, like we did for Assistants API.

[PierreMesure]

Indeed, the openai-node package supports the Responses API since v4.87.0.

[deviprasad97]

Any update on this?

[benjaminsingleton]

I also wanted to express support for natively supporting OpenAI's new Responses API. Without supporting this API, LC won't be able to match the capabilities of the official ChatGPT app when using OpenAI models, which users already expect. For example, OpenAI's o3 model is incredibly capable when it has access to OpenAI's built-in web search tool, but it's not possible using the Chat Completions API.

To respond to earlier commends, the main differences between the Responses API and the Assistants API are:

• Simpler integration: Responses API uses the familiar chat-based schema (messages[] and inline tools) without mandatory threads or polling, significantly easing maintenance.

• Built-in first-party tools: Includes web search, file search, code interpreter, image generation, and more—exactly what ChatGPT uses today, enabling LC to match its capabilities.

• More future-proof: Other AI labs (e.g., Anthropic) are already converging on a similar "response + tools" pattern, aligning with LC's goals of broad model support. Anthropic released their "Messages" API, which is similar to OpenAI's Responses API, particularly in how tools are used.

Here is a table further detailing the differences between the Assistant API and the Responses API:

Because the Responses API keeps the familiar chat schema, LibreChat doesn’t need another vendor‑specific object hierarchy—just the ability to:

• Pass the tools array down to OpenAI when users pick an eligible model.

• Listen for tool_calls in streamed deltas and route them through LibreChat’s existing plug‑in/agent framework.

• My understanding is that is similar to what LibreChat already does for function‑calling models today, so the integration surface is much smaller than the old Assistants work.

In practice, support for this feature could look like:

• Expose a “Built‑in tools” toggle in the parameters / presets.

• When a user enables it, LC simply forwards the chosen tools (e.g., {type:"web_search_preview"}) in the request and treats tool_calls the same way it treats ordinary function calls.

• For users who prefer custom/local tools, LC's agent framework continues to work unchanged—Responses just becomes another transport layer.

[danny-avila]

Without supporting this API, LC won't be able to match the capabilities of the official ChatGPT app when using OpenAI models, which users already expect.

I see this differently - LibreChat's philosophy has always been about providing flexible alternatives rather than replicating official apps exactly. Our native web search feature shows we can deliver comparable functionality through the Chat Completions API.

Your point about implementation is well-taken though. The Responses API's smaller integration surface and alignment with the "response + tools" pattern that providers like xAI, Anthropic, and Google are adopting makes it a sensible addition to support.

I'll be working on this when I get the chance. Thanks for the detailed breakdown!