OpenClaw 2026.5.20 just dropped.
This one fixes some stuff that's been quietly breaking agents for months. Discord voice that follows you into voice channels. Doctor that catches plaintext secrets. xAI you can finally log into without a localhost callback. Windows installs that stop freezing. I'm going to walk you through everything in this update, what it means for your business, and the one setup that makes all of it ten times more powerful.
Your agent walks into the voice channel with you.
Let me start with the Discord voice update because this is the one I've been waiting for.
OpenClaw has always worked from Discord. You message your agent in a text channel, it does the work. Fine.
What was missing — when you moved into a voice channel, your agent stayed behind. Disconnected. The conversation broke.
OpenClaw 5.20 fixes that.
Now your Discord voice sessions follow configured users into voice channels. Your agent moves with you.
- Multi-user handoff — voice session passes between configured users in the same channel.
- Allowed-channel checks — your agent only joins channels you explicitly approve.
- DAVE encryption recovery — session keeps working when Discord rotates encryption keys mid-call.
- Realtime context — IDENTITY, USER, and SOUL profile context now ship with voice sessions by default.
You hop into a voice channel with a teammate. You say, "agent, check the leads from this morning and summarize the top three." Your agent is right there, in the call, doing it. Hands-free.
For community managers, gaming creators, and anyone running Discord ops, this changes the shape of the day.
Doctor now warns you before a leak.
Next big one. Doctor now warns when your openclaw.json file stores plaintext secrets like API keys or sensitive headers.
Before this update, you could have an Anthropic key or an OpenRouter key sitting in plain text in your config. OpenClaw never told you. You found out when somebody else found out.
Now you run openclaw doctor and it flags every plaintext secret before it becomes an incident.
- Model provider API keys sitting in
openclaw.jsonwithout encryption. - Sensitive provider headers — Bearer tokens, custom auth headers, anything secret-shaped.
- Sandbox tool policy conflicts hiding MCP server tools you actually configured.
- Stale provider model config that fails validation after upgrade (gets cleaned by
doctor --fix).
This is the kind of fix most users do not appreciate until it saves them. If you have provider keys sitting in your config right now, run the doctor as soon as you update. You will be glad you did.
No more wondering which model is running.
Status got smarter too. And this fixes something that has been quietly annoying everyone.
When a session is pinned to a different model than your configured default, you used to have no idea why. Maybe a fallback kicked in. Maybe an agent override hit. The status said nothing.
Now the status shows you everything in one place.
- Configured default model — what your config says should run.
- Selected model — what is actually running right now.
- Reason for the difference — fallback, override, manual pin, etc.
- Clear hint — what to do if you want to switch back to the default.
- Direct docs link — clickable, takes you to the exact relevant page.
You run openclaw status and the system tells you exactly which model you are talking to and why. No more guessing why the report came back in the wrong voice.
Logging xAI into a remote server just got easy.
If you have ever tried to authorize xAI on a headless server, you know the pain.
The old login flow needed a localhost browser callback. Which does not exist on a remote machine. So you ended up SSH-tunneling ports just to get logged in. Brutal.
OpenClaw 5.20 adds device-code OAuth for xAI.
- Step 1. Start the login on the remote server. The terminal shows you a short code.
- Step 2. Open xAI on your phone or laptop. Paste the code into the login screen.
- Step 3. You are authorized. Same flow GitHub Copilot uses. Same flow the Anthropic CLI uses.
If you run OpenClaw on a VPS, a Modal sandbox, or a Daytona workspace, this single fix removes one of the most annoying setup blockers of the year.
Windows installs no longer freeze at "Starting setup…"
This one is for every Windows user who tried to install OpenClaw and watched the wizard freeze on a black screen for ten minutes.
The installer now launches as an attached child process. So the wizard's terminal rendering stays clean instead of being eaten by a detached parent. No more visible freeze.
Managed updates also use the Gateway service Node consistently. Before this, if you had multiple Node installations on your machine, OpenClaw would sometimes silently switch which Node it ran on between updates. Hidden source of breakage. Gone.
If Windows install has been the thing stopping you from setting up OpenClaw, today is the day.
The 5 headline features at a glance
Here is what changed, side by side.
If you want all of this in one connected system.
OpenClaw is powerful on its own. It is ten times more powerful when it shares memory with Claude and Hermes.
That is what the Agent OS does.
OpenClaw + Claude + Hermes — one operating system.
If you want to get the most out of these new OpenClaw features — the Discord voice handoff, the doctor warnings, the xAI OAuth — you plug OpenClaw into the Agent OS inside the AI Profit Boardroom.
It is a full operating system I built that connects OpenClaw, Claude, and Hermes into one dashboard. Your agents share one memory. They know your goals. They know your business.
So when you set up the new Discord voice mode, your agent already has full context for everything. When xAI answers a question, it pulls from your shared knowledge base. When OpenClaw ships another fix next week, the whole system gets the benefit automatically.
- The Agent OS zip file — every prompt, every config, the full Obsidian memory setup.
- A 30-day OpenClaw roadmap — exactly how to wire OpenClaw into the Agent OS, step by step.
- Four coaching calls every week with operators running OpenClaw + Claude + Hermes in production right now.
- Daily tutorials walking through every OpenClaw release as it ships.
- 2,800 business owners already running this stack — someone is online 24/7.
- A member map so you can connect with other OpenClaw users near you.
Now let me cover the changes that make your setup more stable.
84 commits in one release. The headlines were 5. The other 79 are fixes nobody is going to tweet about. But these are the ones that quietly keep your agents alive.
Eight of the most important ones, in plain English:
- Subagent handoffs no longer drop. Research-agent → writer-agent results are preserved. No more lost work.
- Cron jobs survive upgrades. Old format files are no longer wiped during an update.
- Anthropic routing now respects Claude CLI auth. No more double-billing when you use shorthand model refs.
- Codex encrypted reasoning replay is provenance-bound. Stale transcripts no longer crash your subagents at 2 AM.
- Memory search properly closes local embedding providers when searches time out. No more zombie models eating RAM.
- Plugin hooks get a 30-second default timeout. A hung plugin can no longer block context compaction.
- Credential symlink protection for Telegram, LINE, Zalo, IRC, Nextcloud Talk. Symlinked credential files are refused.
- Gateway version is now exposed in status JSON. Scripts can finally read the running version programmatically.
And that is 8 out of 79. Every single one came from a real failure someone hit in production. None of it ships on platforms that aren't taking this seriously.
This is a stable release. Backup first.
Should you update? My advice hasn't changed.
This is a stable release. It went through validation. The community reaction in the first three hours has been positive — 33,900 views on the announcement post, and the replies are people saying the Discord voice mode actually works, the doctor is catching real secrets, and the Windows install finally runs.
But still — if your setup works today, back it up first. Run openclaw backup create before you touch anything. Know your current version. Test after updating. If something breaks, roll back.
The four-command routine:
$ openclaw backup create $ openclaw update $ openclaw doctor $ openclaw status
That's it. Ten minutes. You're current.
The people who figure out AI agents now, while the tools are evolving fast, are going to be way ahead when everything settles. Every update you learn. Every workflow you build. Every problem you solve. It all compounds.
Make your OpenClaw setup actually save you time every day.
OpenClaw is great. OpenClaw inside the Agent OS, with Claude and Hermes sharing memory, is a different category of thing entirely.
Plug OpenClaw into the Agent OS. Save hours every week.
If you want your OpenClaw setup to actually save you time every day — not just be another tool you check sometimes — grab the Agent OS inside the AI Profit Boardroom.
It turns OpenClaw, Claude, and Hermes into one system with shared memory, shared context, and one dashboard you control. Your agents understand your business. They remember everything.
Every new OpenClaw update — like the Discord voice mode and the xAI OAuth that just shipped — makes the whole system more powerful automatically. You do not rebuild. You plug in.
- The Agent OS zip file built in one session — every prompt, every config, the Obsidian memory setup.
- A 30-day OpenClaw roadmap walking you through OpenClaw + Claude + Hermes step by step.
- Coaching calls every week where I set it up live with you. Bring your questions.
- Daily tutorials walking through every OpenClaw release as it ships — so you stop reading patch notes at midnight.
- 2,800 members already running this stack. Someone is online 24/7.
- A member map so you can connect with other OpenClaw users near you.
- A prompt library built around OpenClaw workflows — Discord ops, support agents, content pipelines, lead routing.
If you've been hitting walls with OpenClaw — failed workflows, broken integrations, secrets in plaintext, agents that work for two weeks then stop — this is where to fix it.
Run openclaw backup create first. Decide second. I'll see you in the next one.