OpenClaw Hid Its Best Feature. Maybe on Purpose.

OpenClaw shipped its third release in three days on March 25, and buried the lede inside the changelog. Buried it so deep that the blog post announcing 3.24 spent most of its pixels on Teams SDK overhauls and skill management UI before mentioning that the Gateway now exposes OpenAI-compatible /v1/models and /v1/embeddings endpoints. That is the actual story.

The short version: if you have existing tooling built for OpenAI's API, you can now point it at an OpenClaw Gateway and it should work. Not "use our SDK" — just connect. The /v1/models endpoint lists available models, /v1/embeddings generates them, and explicit model overrides forward through /v1/chat/completions and /v1/responses. GitHub's v2026.3.24 release notes have the full changelog; the OpenAI compat section is near the bottom, as if the team was quietly hoping nobody would notice.

This matters because the integration tax on agent frameworks has been real. Getting OpenAI's client libraries to talk to anything other than OpenAI's API required custom adapter code, wrapper libraries, or explicit SDK support. OpenClaw's new endpoints eliminate that friction for anyone already running an OpenAI-compatible toolchain — RAG stacks, agent frameworks using LangChain-style adapters, evaluation harnesses built against the OpenAI API spec. The MIT-licensed, self-hosted agent platform just became accessible to anyone who already knows how to call an LLM API. Whether that's a deliberate strategy or a side effect of how the team architected the Gateway is unclear; the effect is the same either way.

Tencent integrated ClawBot into WeChat on March 22 — one day before OpenClaw 3.22 shipped. That puts ClawBot in front of over a billion monthly active users. Tencent was working against 3.21 and shipped anyway, a moving-target integration that speaks to the commercial pressure on both sides. China's Ministry of Industry and Information Technology had flagged concerns on February 5 — saying default or poorly configured deployments create serious exposure risks. Tencent's integration went live twelve days after that warning. The attack surface the Ministry flagged is the same surface Tencent bet its commercial agent product on.

The release that preceded 3.24 explains why the team was moving fast. OpenClaw 3.23, released March 24, was the emergency hotfix — the team pushed another release the day after 3.22 landed because the npm tarball for 3.22 had shipped broken. The Dashboard returned 503 errors and the WhatsApp plugin had simply vanished from the published package. The 3.23 changelog lists 40+ stability fixes and eight patches specifically to the auth and credential system, including one where live gateway auth-profile writes were silently reverting saved credentials back to stale in-memory values. "The release that makes sure the patient survives the surgery," the blog put it. "One day between releases, because some of these could not wait."

That followed OpenClaw 3.22 on March 23, which had been the largest release in months after a nine-day gap: 12 breaking changes, 30+ security hardening patches, 100+ stability fixes, seven new model providers, cold starts dropping from minutes to seconds, and ClawHub becoming the default plugin store. The nine-day silence before 3.22 made the changelog land harder — when the project that averages a release every two to three days goes quiet for nine, something big is cooking. The architecture overhaul was the cooking.

Three releases in three calendar days is an unusual cadence, even for a project that ships frequently. The 3.24 release has no breaking changes listed — a clean patch on top of an emergency patch on top of a major architecture overhaul. The security fix in 3.24 closes a mediaUrl/fileUrl alias bypass that could have let outbound tool and message actions escape media-root restrictions. Nothing dramatic, but specific. The kind of thing that gets found when a lot of code is changing fast.

The most immediate user-facing change in 3.24 is the /tools command: agents now see only the tools they can actually use right now, with a compact default view and a live "Available Right Now" section in the Control UI. One-click install recipes for seven bundled skills — coding-agent, gh-issues, openai-whisper-api, session-logs, tmux, trello, weather — land in the same release. Whether that matters depends on how much time you've spent debugging skill availability in a running agent session. If you've done that work, the answer is obvious. If you haven't, the /tools change will feel like a nice-to-have until it saves you two hours.

For teams already running OpenAI-shaped tooling, what 3.24 changes is concrete. The /v1/models and /v1/embeddings endpoints mean your existing RAG stack or LangChain-style adapter works without rewrites. Model overrides through /v1/chat/completions let you swap the underlying provider without touching client code. The integration tax that made custom adapter code worth writing has just been removed — what used to be a migration project is now a config change.