OpenAI Models in OpenClaw, Done Right
10 hours ago
- #Agent Platform
- #Codex Integration
- #OpenClaw
- OpenClaw now uses Codex app-server as the default runtime for OpenAI agent turns, improving integration and reducing friction.
- Codex manages the low-level OpenAI loop: native thread state, tool continuation, compaction, code mode, and dynamic tool search.
- OpenClaw handles product-specific features: channels, persona, memory, sessions, cron, media, browser, gateway, and OpenClaw tools.
- Visible replies are intentional via the OpenClaw message tool, distinguishing internal reasoning from user-facing communication.
- Heartbeat turns use structured state (e.g., heartbeat_respond) for explicit outcomes, replacing sentinel text.
- Dynamic tool loading reduces prompt bloat, allowing models to search and load tool schemas on demand.
- Auth integrates with ChatGPT subscriptions; OpenClaw routes OpenAI models through Codex by default.
- State isolation ensures OpenClaw agents have separate Codex homes, preventing leaks with personal CLI setups.
- Guardrails include reviewed approval modes, leveraging Codex safety while maintaining OpenClaw policy control.
- Benefits extend to all models: lessons from Codex improve OpenClaw's harness for broader multi-model support.