OpenAI Agents SDK Compatibility

Status: current Date: 2026-05-20

This project follows the same core agent-loop shape as the OpenAI Agents SDK, but it is not a drop-in reimplementation of the whole SDK.

Similarities

Implemented locally:

• Agent loop: model response, tool call, tool result, repeat, final answer.
• Tool schemas.
• Tool execution owned by application code.
• Session-based multi-turn chat.
• Local memory store.
• Policy checks around tools and workspace access.
• Trace logging.
• HTTP /chat and /run service shape.
• Runtime provider/model selection through /registry/models and /models/select.
• Ollama model adapter for local chat.
• OpenAI-compatible chat-completions adapter for hosted or compatible APIs.
• Embedded browser voice page using /chat.

This matches the conceptual flow described in OpenAI’s running-agents docs: model call, inspect output, execute tool calls, continue, or return final answer.

Not Implemented

This repo does not currently implement all OpenAI Agents SDK methods or features.

Missing or only represented as local analogues:

• Official Agent, Runner, and SDK result types.
• Official OpenAI Agents SDK model adapter. A local OpenAI-compatible chat completions adapter exists, but it is not the SDK runtime.
• OpenAI Conversations API integration.
• previous_response_id continuation.
• Streaming agent runs.
• Agent handoffs/orchestration across specialists.
• Hosted OpenAI tools.
• MCP tool registration.
• Human-in-the-loop approval pause/resume.
• OpenAI tracing/evals integration.
• Sandbox agents.
• ChatKit/frontend widgets.
• OpenAI realtime voice agents. A browser voice page exists locally; see Voice Agents.

Recommendation

Use this repo as the provider-neutral backend spine:

chat/session service
-> model selection interface
-> model adapter
-> agent loop
-> policy
-> tools
-> memory
-> traces/evals

Then add an OpenAI Agents SDK adapter if we want the official SDK runtime and hosted capabilities.