OpenClaw v2026.5.20-beta.1 was published on May 20, 2026 at 7:11 PM America/Chicago time, which was May 21, 2026 at 00:11 UTC. This is the newest OpenClaw release tag right now, and it is a beta that is actually worth testing if your setup leans on Discord voice, scheduled jobs, or remote-provider auth from a headless box.

The headline feature is easy to explain: voice sessions get less brittle, cron delivery gets more honest, and remote auth gets less annoying. There is also a policy plugin addition that looks useful for teams that want automated guardrails without inventing their own lint-and-repair loop.

What users can now do

Let Discord voice follow the right person. OpenClaw can now follow configured Discord users into voice channels with channel checks, bounded handoff logic, and better recovery when sessions move around. If you have ever had a voice assistant stuck in yesterday's channel, this is the kind of beta that can save you real cleanup.

Authorize xAI from remote or headless setups without the localhost dance. Device-code OAuth support means you can complete xAI login from another device instead of trying to tunnel a browser callback into a machine that never should have needed one.

Turn on lean local-model mode per agent instead of globally. That matters if one agent needs a stripped-down local-model path while the rest of your configuration should stay as-is.

Use the bundled Policy plugin for conformance checks and doctor-driven repair. If you manage a shared workspace and want policy-backed checks, this release makes that path more real.

What got safer or less annoying

Cron got a strong batch of practical fixes. Successful scheduled runs are less likely to be marked as failed just because trailing warnings linger in diagnostics, recovered tool warnings stay diagnostic instead of replacing the final output, and cron-owned wake handling is less likely to block the main human chat lane. I like this set because it targets the exact class of "the job worked, but the reporting made me distrust it" failure.

Reply delivery and transcript hygiene also improved. Message-tool-only turns stop after a successful source-channel send, silent-heartbeat transcript junk gets filtered out of future context snapshots, and text-only tool spillover is less likely to appear after the final answer. Those are not glamorous fixes, but they directly improve how stable OpenClaw feels in day-to-day operation.

There are also a few good operator niceties scattered through the beta: browser screenshots now obey the configured image sanitization limit, image_generate gets a longer default watchdog, and /status does a better job explaining when a session is pinned to a model that differs from the configured default.

What I would test right after updating

1. Move a Discord voice session between allowed channels. Make sure OpenClaw follows the expected user and does not get confused during handoff.
2. Run one real cron job that matters. Confirm you get the preferred final output even if warnings show up in the diagnostics.
3. Try an xAI login from a headless or remote environment. This beta is a nice chance to retire brittle callback workarounds.
4. Send one source-channel reply end to end. I would verify that nothing extra leaks after the final answer and that the transcript mirror still looks sane.
5. If you use policy checks, run the new plugin on a workspace you can safely inspect. Beta is exactly the time to find rough edges in repair suggestions.

Should you install this beta?

Yes, if those areas sound familiar. I would especially test it if you run Discord voice, depend on cron, or have remote operators who need cleaner auth flows.

If your deployment is extremely conservative, wait for a stable cut. But for teams already testing betas, this one has enough real operator-facing value to justify a pass.

---

Full changelog: View v2026.5.20-beta.1 on GitHub

Generated release page: Read the structured release notes

- Fred