We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Feb 4·5 min read
Unlocking the Codex harness: how we built the App Server
TL;DR
Integrating Codex, an AI code generation model, into production applications required a standardized interface for agent communication. OpenAI built the Codex App Server, a bidirectional JSON-RPC API (a lightweight protocol for remote procedure calls) enabling streaming output, tool integration, user approvals, and code diff visualization.
✦ Why It Matters
You can now embed Codex agents in production with standardized APIs, real-time control, and safety approvals built in.
Key Takeaways
- Use Codex App Server's JSON-RPC API to standardize agent-application communication in production.
- Implement streaming progress updates to provide real-time visibility into Codex agent execution.
- Integrate approval workflows to require human sign-off before Codex executes code changes.
- Leverage diff visualization to show code modifications before deployment, improving safety and transparency.
Full Summary
Codex is OpenAI's AI model trained to generate and understand code. Previously, deploying Codex in production applications lacked a standardized communication layer between the AI agent and host applications.
OpenAI developed the Codex App Server, a bidirectional JSON-RPC API that establishes a two-way communication protocol between Codex and external systems. The server supports streaming progress updates (allowing real-time visibility into agent execution), tool use (enabling Codex to call external functions), approval workflows (requiring human sign-off on actions), and diff visualization (showing code changes before execution).
This architecture decouples the Codex agent from application-specific logic, making it reusable across different platforms. Engineers can now embed Codex with built-in safety mechanisms and user control, reducing integration complexity and enabling broader adoption of AI-assisted coding.