Open Responses - The NEW Standard API for Open Models

Quick Overview

Open Responses is introduced as a new open-source specification and ecosystem, based on the OpenAI Responses API, designed to provide an interoperable, unified interface for building multi-provider LLM applications, enabling features like consistent streaming, tool use, and reasoning traces across different model providers like OpenAI, Anthropic, Hugging Face, OpenRouter, Vercel, LM Studio, Ollama, and VLLM.

Key Points: Open Responses is an open-source specification and ecosystem built around the OpenAI Responses API to create a unified, multi-provider interface for LLMs. The specification supports key features like consistent streaming events, tool invocation patterns (both externally-hosted and internally-hosted tools), and exposing reasoning traces. Key community adopters and supporters highlighted include OpenAI, Anthropic, Hugging Face, OpenRouter, Vercel, LM Studio, Ollama, and VLLM. The framework introduces a state machine for Items with three basic statuses: inprogress, incomplete, and completed, ensuring controlled response flow. Reasoning items expose the model's internal thought process via content (raw/encrypted) or summary, allowing developers to see model steps or opt for privacy. The specification standardizes tool calling, allowing models to use tools (like 'getcurrentweather') defined in a provider-agnostic manner. The live demo showcases compatibility with Ollama running locally, successfully executing basic calls, event-based streaming, tool calling, and reasoning traces.

Context: The video introduces Open Responses, a new initiative aiming to standardize the API layer for interacting with various Large Language Models (LLMs). This effort seeks to solve the fragmentation problem where different providers (like OpenAI, Google's Gemini, Anthropic, etc.) use slightly different API schemas for common features such as tool use, streaming, and reasoning.

Detailed Analysis

Open Responses establishes a shared, open specification for LLM interfaces, inheriting structure from the OpenAI Responses API to ensure interoperability across providers. Its design goals include being multi-provider by default with one schema, being friendly to real-world agentic workflows with consistent streaming and tool patterns, and being extensible without fragmentation. The specification defines core concepts like the Agentic Loop, where models iteratively perceive, reason, act through tools, and reflect on outcomes. Items are the fundamental unit of context, tracked through a state machine (inprogress, incomplete, completed). It standardizes tool use, distinguishing between externally-hosted tools (where control yields back to the developer after execution) and internally-hosted tools (where the model executes the tool and returns results without yielding control). Reasoning is exposed via specific item types, providing raw trace, encrypted content, or summaries. The presenter demonstrates successful execution using Ollama locally, showing basic API calls, event-based streaming, tool calling (using a weather function), and reasoning traces, proving the specification's immediate utility across different model types and local/cloud inference solutions.