Codex App Server: custom clients and codex exec comparison

Dominik Kundel from OpenAI explains how developers can build a custom AI agent interface on the Codex harness through App Server. He describes a JSON-RPC style protocol that carries client requests and server events, then uses an inspector to demonstrate thread creation, configuration reads and streamed responses.

Kundel says the harness and App Server are open source under Apache 2. He describes support for other providers with a Responses-compatible API and convenience settings for LM Studio and Ollama, relevant to developers who want to run models locally. The talk does not demonstrate an offline setup.

Customization covers developer instructions appended to the Codex system prompt, dynamic tools with deferred discovery, and per-thread model and sandbox settings. Kundel also explains ChatGPT sign-in, which lets users draw on their own Codex usage limits.

App Server ships in the Codex CLI binary and supports STDIO or WebSockets. Kundel recommends bundling a specific version because the protocol changes between releases; developers can generate matching TypeScript bindings or JSON schema. He introduces the Python SDK for lifecycle management and demonstrates game tools inside Doom. For scripting and CI/CD, he recommends codex exec; for interactive clients or concurrent threads, he favors App Server.