OpenClaw shipped a release on May 15 that changes how its agents interact with OpenAI models. The Codex app-server harness is now the default runtime for all openai/gpt-* agent turns, replacing a previous approach where OpenClaw drove the model loop itself.
“Your ChatGPT subscription can now power an OpenClaw agent that feels much closer to the model it is built on,” OpenAI’s Nik Pash wrote in a blog post announcing the change. OpenClaw founder Peter Steinberger, who joined OpenAI last year, said the team has been working “really hard on performance, reliability, security, and stability,” according to The Verge.
What Changed Architecturally
The release draws a cleaner boundary between what OpenClaw owns and what Codex owns.
Codex now handles the inner model loop: native thread state, tool continuation, compaction, code mode, and dynamic tool search. OpenClaw retains control of the product layer: channels, persona, memory, sessions, cron, media, browser automation, gateway, and OpenClaw-specific tools.
The blog post describes the prior setup as OpenClaw “translating between its own harness and the runtime OpenAI is actively building for agentic work.” Removing that translation layer means less duplication. The model no longer has to choose between duplicated workspace tools, and OpenClaw no longer has to pretend Codex-native tools (read, edit, patch, exec, process, planning) are generic plugin tools.
Dynamic Tool Loading
The most significant practical change is how tools are surfaced. OpenClaw agents can carry dozens of tool schemas: messaging, sessions, media, cron, browser, nodes, gateway controls, web search, MCP servers, plugin tools, and channel-specific actions. Previously, all schemas were loaded into the initial prompt.
With the Codex harness, OpenClaw now passes its capabilities as searchable dynamic tools. Codex discovers the right tool on demand through native tool search, keeping initial context smaller and reducing the chance of the model picking the wrong tool.
The blog post notes this pattern is already flowing back to non-OpenAI models. An experimental feature called PI Tool Search gives non-Codex runs a compact search-describe-call surface instead of preloading every tool schema. It is not on by default yet, but the direction signals that the Codex integration is producing architectural improvements for the entire platform.
Reply Delivery and Heartbeats
Codex-backed turns now use the OpenClaw message tool for visible replies rather than relying on the final assistant string becoming the visible message. The distinction: internal reasoning stays private, visible replies are intentional, and quiet turns (like heartbeats) remain actually quiet.
Heartbeats shifted from sentinel text parsing to structured tool calls. Instead of emitting “HEARTBEAT_OK” as a string for OpenClaw to interpret, the agent calls heartbeat_respond with explicit outcomes: “nothing to report,” “notify the user,” or “schedule a follow-up.”
Platform Interoperability Signal
The release is notable beyond its technical specifics. It validates a decoupled architecture for AI agents: the inference provider (OpenAI/ChatGPT) and the orchestration platform (OpenClaw) operate as separate layers with a clean interface between them. Developers can use their existing ChatGPT subscription without paying twice, while choosing their agent framework independently.
This pattern contrasts with vertically integrated approaches where the LLM provider also owns the agent runtime, the tool execution, and the delivery surface. OpenClaw’s bet is that the product layer around the model turn, not the model turn itself, is where agent platforms should compete.
Setup is available through openclaw onboard or openclaw models auth login --provider openai, according to the OpenClaw documentation.