OpenClaw adds ChatGPT subscription sign-in for hosted coding-agent sessions

• sama's announcement says OpenClaw now lets users sign in with a ChatGPT account and use an existing subscription inside the harness.
• The new path sits on OpenClaw's openai-codex/* route, which the agent runtimes docs describe as ChatGPT or Codex subscription OAuth running through the normal PI runner, not the API key path.
• The feature landed as part of a fast-moving Codex integration cycle: the 2026.4.27 release notes changed how ChatGPT subscription app-server runs inherit auth, while PR #75111 fixed broken OAuth transport and preserved account plan context.
• The sanctioned route still has edges. Issue #73963 shows a 400-breaking metadata bug on the ChatGPT subscription path, and Issue #74451 shows openai-codex/gpt-5.4-mini was unsupported for ChatGPT-backed Codex accounts.
• The timing sharpened the contrast with Anthropic discourse: while theo's terminal screenshot and the main HN thread focused on Claude Code refusing or overbilling around OpenClaw mentions, OpenClaw was opening a first-party subscription login path for the same competing harness.

You can read the v2026.4.27 release notes, the agent runtimes docs, and the OAuth docs. The interesting bit is how much plumbing was required just to make a consumer ChatGPT subscription behave like a usable third-party coding backend: PR #75111 had to preserve OAuth injection and account metadata, while Issue #73963 and Issue #74451 show the path was still brittle days earlier.

The user-facing change was simple. OpenClaw users can now authenticate with a ChatGPT account and spend the subscription they already have inside OpenClaw's hosted coding sessions.

That matters mostly because it turns a previously fiddly backend route into an explicit product surface. The announcement came from Sam Altman, then got echoed by OpenAI product lead Nick Turley in nickaturley's follow-up, which is about as official as third-party harness support gets.

OpenClaw's docs split the OpenAI integration into separate surfaces:

• openai-codex/* uses ChatGPT or Codex subscription OAuth through the normal PI runner, per the agent runtimes docs
• agentRuntime.id: "codex" runs the bundled native Codex app-server harness, also documented in agent runtimes
• openai/* keeps using the OpenAI Platform API key route unless a runtime override is set, again per agent runtimes

The implementation work was more involved than the launch tweet suggests. The 2026.4.27 release notes say ChatGPT subscription app-server runs should stop inheriting CODEX_API_KEY or OPENAI_API_KEY, and only fall back to those keys when no Codex account is available.

Then PR #75111 fixed the transport layer itself: preserve the embedded OAuth injection stream, strip fields the native ChatGPT Codex backend rejects, and persist OAuth account metadata so refreshed profiles keep account and plan context.

The docs are unusually explicit on policy. In the OAuth docs, OpenClaw says OpenAI Codex OAuth is clearly supported for outside tools like OpenClaw, while Anthropic subscription auth inside OpenClaw was described separately and attributed to conversations with Anthropic employees.

The subscription path shipped with real limits and real bug reports.

• Issue #73963: openai-codex/* requests to chatgpt.com/backend-api/codex/responses were rejected with HTTP 400 because OpenClaw attached a metadata field the backend did not accept
• PR #75111: another fix preserved the OAuth bearer on embedded runner requests after some paths were returning 401 status code (no body)
• Issue #74451: ChatGPT-backed Codex accounts could still hard-fail if a session switched to openai-codex/gpt-5.4-mini
• The 2026.4.29 release notes added another guard so stale configured model rows could not bypass that capability block

So the story here is not just "OAuth exists now." It is that OpenClaw spent the week hardening a consumer subscription route into something agents can actually sit on.

The competitive backdrop was unusually blunt. In theo's post, a Claude Code call in an otherwise empty repo hit an extra-usage error after a commit message included openclaw.inbound_meta.v1; the HN thread amplified similar reports and user frustration.

That is why Altman's "happy lobstering" line landed. OpenClaw was not merely proxying API keys. It was advertising a vendor-approved way to point a mainstream ChatGPT subscription at a rival coding harness while the competing harness was being accused of refusal and billing heuristics around the word OpenClaw.