This beta is for people who use OpenClaw all day and are tired of the small stuff getting in the way. Voice replies got the biggest upgrade. Browser automation got safer. Plugin startup got less weird.

OpenClaw published v2026.4.25-beta.4 on April 26, 2026. The changelog is broad, but the practical story is easier to tell: you can now make voice replies part of your normal workflow instead of a side experiment, and the parts that boot, inspect, and update your system are less likely to surprise you.

Because this is a beta, I would still test it on the setup you actually care about before rolling it everywhere. But if you use TTS, browser automation, or a pile of plugins, this is worth your time now.

Voice replies are finally something you can leave on

The headline feature is the TTS refresh. /tts latest can read the latest reply aloud. /tts chat on and /tts chat off let you control auto-TTS at the chat level instead of treating voice as a global all-or-nothing setting. Per-agent and per-account overrides are in the same release, which matters if one bot should sound different from another or one account needs different defaults.

This fixes a real annoyance. Before this, TTS could work, but it often felt like a demo path. You had to think about config more than you should. Now it is much closer to a normal messaging feature.

/tts latest
/tts chat on
/tts status

Azure Speech joins the bundled providers, and the release also adds coverage for Xiaomi, Inworld, Volcengine, Local CLI voices, and ElevenLabs v3. If you have been waiting for a voice stack that matches the rest of your setup, there is a good chance your preferred path is here now.

What users can do now

• Read the last answer aloud on demand. That is the new /tts latest path.
• Turn chat-level auto-TTS on without changing every other chat. Use /tts chat on where it helps and leave it off elsewhere.
• Give different agents different voices. The new overrides finally make that configuration sane.
• Use Azure Speech as a bundled TTS provider. No sidecar glue job needed.
• Push browser work a little harder on slow machines. You can raise Chrome discovery and readiness timeouts, and there is now a one-shot headless browser start for local runs.

Browser automation should waste less of your time

A lot of this release is about reducing ambiguity. Agent responses now report safer tab URLs. Role snapshots have a CDP-native fallback and understand iframes better. The browser doctor got deeper probes for slow hosts. On paper that sounds technical. In practice it means fewer moments where the browser was open, the page looked fine, and the agent still felt half blind.

If you run browser automation on a Raspberry Pi or another slow box, the readiness timeout changes matter right away. If you drive the browser from the CLI, openclaw browser start --headless gives you a clean one-shot launch without rewriting stored browser config.

Plugin startup and updates got more boring in the best way

Plugin registry work takes up a huge chunk of this changelog. Good. Startup and install paths now lean on the persisted registry instead of broad manifest scans. That should make plugin discovery, repair, update metadata, and provider ownership much more deterministic.

If you have ever watched a machine with a lot of plugins boot slowly or fail in a way that was hard to reproduce, this is the kind of release you want. There are also a lot of install and repair fixes around Windows, macOS, Linux, Docker, runtime dependencies, and service restarts. None of that is glamorous. All of it matters when the machine has to come back up cleanly at 2 a.m.

Operators get better visibility too

OpenTelemetry coverage expands across model calls, token usage, tool loops, harness runs, exec processes, outbound delivery, and memory pressure. There is also a bundled Prometheus diagnostics plugin and cleaner signal-specific OTLP endpoint control.

You do not need observability to use OpenClaw. But if you run it for a team, or you are the person who gets paged when it goes sideways, this release gives you better traces without making you duct-tape everything together first.

What got safer or less annoying

• Install and upgrade paths got harder to knock off course. The release adds more checks around mixed-version gateway verification, LaunchAgent token rotation, and runtime dependency repair.
• Voice channel routing is more precise. Discord can now override the LLM for voice responses without dragging STT and TTS settings along with it.
• The first-run experience is less fragile. Crestodian setup picked up a TUI flow, progress indicators, a local planner fallback, and a context mode selector.
• Google Meet exports are no longer a one-off hack. Attendance export workflows now have manifests, dry-run previews, and tool parity.

What I would test right after updating

1. Run /tts latest in a real chat. Then turn /tts chat on and make sure the next reply behaves the way you expect.
2. If you use a custom voice provider, switch voices on purpose. Check that your per-agent or per-account override wins where it should.
3. Run one browser task that used to be flaky. Pick a page with an iframe or a slow load and see whether the agent keeps the right tab context now.
4. Restart the gateway on the box you actually depend on. Watch plugin startup, provider discovery, and any runtime dependency repair path.
5. If you collect telemetry, confirm traces and metrics still land where you expect. This release touches that plumbing heavily.

Should you install this beta now?

Yes if voice replies are part of your daily use, if browser automation is on the critical path, or if plugin startup has been one of those problems you keep half-fixing.

Wait for stable if your OpenClaw box is quiet, boring, and already doing exactly what you need. This release looks promising, but it is still a beta, and the people who get the most value out of it are the ones already hitting the edges it tries to smooth out.

The best part of this release is not one flashy feature. It is that more of OpenClaw now behaves like a tool you can leave running without babysitting it.

---

Full changelog: View v2026.4.25-beta.4 on GitHub

Generated release page: Read the structured release notes

- Fred