The open-source AI assistant platform OpenClaw has handed over OpenAI model inference, tool calling, and context management to the Codex runtime in its latest release, v2026.5.12. Previously, all of this was driven by self-developed code. Now, the OpenAI side is fully handed to OpenAI's own runtime, with the reasoning that Codex is the best host for the OpenAI model loop.
The direct motivation is tool bloat. OpenClaw's Agent can mount a ton of tools — browsers, MCP servers, calendars, plugins — and stuffing them all into the initial prompt burns tokens and makes mis-selection easy. Codex's native dynamic tool loading lets the model retrieve schemas on demand, rather than reading everything upfront. OpenClaw is already porting this approach to its own PI runtime, so Claude, Gemini, and DeepSeek can benefit too.
Reply paths have tightened. After switching to Codex, visible messages must be actively sent via a message tool; internal reasoning stays hidden by default, so the model's intermediate thoughts won't accidentally end up in a group chat. Users just log in with their ChatGPT subscription — no separate API key needed — and each Agent's Codex environment is isolated.
The same version also moved WhatsApp, Slack, Amazon Bedrock, and other plugins out of the core package into standalone installs, and hardened security on the gateway, sandbox, and node pairing fronts.