OpenClaw 2026.5.18: Grok OAuth, realtime Android voice, and production polish
OpenClaw 2026.5.18 is a plumbing-heavy release: Grok OAuth fixes, realtime Android Talk Mode, Telegram topic delivery repairs, browser dialog handling, Proxyline 0.3.3, Node.js 22.19+, and better startup tracing.
OpenClaw 2026.5.18 is live, and this one is not about a single shiny demo feature. It is the kind of release that makes a 24/7 agent stack feel less fragile: authentication fixes, better voice coverage, Telegram delivery repairs, browser dialog handling, stricter proxy work, and cleaner startup traces.
That matters because production-grade AI infrastructure is mostly won in the boring places: auth refreshes that do not randomly drop, message delivery that lands in the right thread, browsers that do not get stuck on popups, and startup traces that tell you where time actually went.
Key takeaways
- xAI/Grok OAuth is more reliable. The release includes OAuth and sidecar auth fixes for xAI/Grok, including token-flow compatibility and sidecar login work.
- Android Talk Mode is now realtime. Android Talk Mode moved to realtime Gateway relay voice sessions with streaming mic input, realtime audio playback, tool-result bridging, and on-screen transcripts.
- Telegram forum-topic delivery got safer. Generated media and follow-up delivery now preserve topic IDs more reliably instead of falling back into the wrong place.
- Browser dialogs are no longer invisible blockers. Browser snapshots can surface modal dialogs, and agents can answer pending dialogs instead of sitting stuck.
- The release tightens infrastructure. Proxyline moves to 0.3.3, Node.js minimum support moves to 22.19, Docker/Podman builds get a cleaner package hook, and Gateway/ACPX startup tracing is more attributable.
What changed
| Area | What changed | Why it matters |
|---|---|---|
| Grok / xAI auth | OAuth and sidecar auth fixes, including PKCE/token endpoint compatibility and sidecar login cleanup | Fewer random auth drops in long-running agent rotations |
| Android voice | Realtime Talk Mode via Gateway relay sessions | Your phone becomes a live voice interface for your agent, not just a slow voice-to-text wrapper |
| Telegram | Media completion and forum-topic delivery fixes | Agents in topic-based Telegram groups can reply where the conversation actually started |
| Browser automation | Modal dialogs are visible in snapshots and answerable | Cookie banners, confirms, and popups are less likely to freeze unattended browsing |
| Security/proxy | Proxyline updated to 0.3.3, with additional proxy work in the release | The network-control layer keeps moving forward with tighter boundaries |
| Runtime requirements | Minimum Node.js 22 line raised to 22.19 | Operators should update Node before updating OpenClaw |
| Containers | OPENCLAW_IMAGE_APT_PACKAGES added for Docker/Podman builds | Cleaner custom image builds across container runtimes |
| Startup visibility | Gateway and ACPX restart traces attribute startup probe, config, runtime, and resource-count costs | When startup is slow, you can see where the time went |
Why it matters
If Grok is part of your model rotation, authentication stability is not cosmetic. A model that works only until a refresh token silently fails is not production-ready. The 2026.5.18 xAI/Grok fixes are aimed at that exact operational pain: keeping subscription-backed and OAuth-backed flows steadier during real sessions.
The realtime Android Talk Mode change is just as important. OpenClaw already had multiple voice paths across devices and channels. This release makes Android feel much closer to a live agent call: streaming microphone input, realtime audio playback, tool results bridged into the conversation, and transcripts on screen.
Telegram fixes are another practical win. If you run agents inside Telegram groups with forum topics enabled, thread routing matters. A correct answer in the wrong topic is still operational noise. The release specifically calls out preserving topic IDs across generated media completions, audio preflight, skipped-message hook contexts, and related handoffs.
For more practical agent workflows, see LinkLoot's guide hub for AI agent tools and AI workflow automation.
The small fix that saves the night
Browser dialog handling sounds tiny until you have an agent running unattended at 3 a.m. and a cookie consent dialog blocks the whole job.
OpenClaw 2026.5.18 surfaces pending and recently handled modal dialogs in browser snapshots, reports when an action is blocked by a dialog, and adds a way to answer pending dialogs by dialog ID. That is not flashy. It is exactly the sort of plumbing that keeps browser automations alive.
Operator notes before upgrading
- Update Node.js first. This release raises the supported Node.js 22 minimum to 22.19.
- Check Grok/xAI login flows after updating. The auth fixes are the headline, but production operators should still verify their own sidecar login state.
- If you run Telegram forum topics, test topic replies. Confirm generated media, replies, and scheduled/proactive flows land in the expected topic.
- If you build containers, migrate toward
OPENCLAW_IMAGE_APT_PACKAGES. The older Docker-specific build arg remains as a legacy fallback, but the new name is runtime-neutral. - Use restart traces when startup feels slow. Gateway and ACPX startup costs are now easier to attribute.
Source check
The primary source is the official GitHub release for v2026.5.18, published on May 18, 2026. The release notes list the xAI/Grok auth fixes, Android realtime Talk Mode, Telegram delivery repairs, browser dialog handling, Proxyline 0.3.3, Node.js 22.19 minimum, Docker/Podman build arg, and Gateway/ACPX tracing changes.
The public announcement on X frames the release as a week of "polish, plumbing, and fewer papercuts." GitHub also currently shows the OpenClaw repository above 373,000 stars, which matches the scale called out in the announcement context.
Bottom line
OpenClaw 2026.5.18 is not a hype release. It is a reliability release.
Grok auth becomes less brittle. Android voice becomes more realtime. Telegram topic routing becomes less annoying. Browser automations become harder to wedge. Startup debugging becomes more transparent. Container builds become cleaner. The proxy layer moves forward.
That is what production AI agent infrastructure looks like: not just new capabilities, but fewer ways for the existing ones to fail.
For operators, the xAI/Grok OAuth and sidecar auth fixes are probably the most important because authentication failures can break long-running model rotations.