It’s tempting. One agent, one machine, a whole team in a single thread, with work handed over along with its context. The new version of OpenClaw lures you with the ability to have multiple users in one installation, in one thread. Limits? There are some!
OpenClaw is an open-source personal agent that you run on your own hardware and talk to through Telegram, Slack, WhatsApp, or a web chat. The project is backed by the OpenClaw Foundation. On August 30, version 2026.8.1 came out, which the team itself calls “the second one” (2.0), and it is the largest release in the project’s history.
Seven weeks of silence
The numbers from the announcement: 933 contributors, 569 of them first-timers, and over 16,000 pull requests. That is roughly half of everything the project has ever received. Before that, they shipped 106 versions in 230 days, so almost every other day.
Version 2.0 was not in the plan. The team started simplifying installation and rewriting the browser app, and once they did it properly, they had to touch everything else. This admission is right at the start of the announcement, and I like it. A big release is usually accompanied by a press release; here you get a factual description of how the work escalated.
What has really improved
Installation uses what you already have on the machine. The wizard looks for a Codex, ChatGPT, or Claude CLI login, picks up an API key, and finds models you’ve downloaded in Ollama and LM Studio. And most importantly: it verifies that the chosen model actually responds before saving it. The old version let you get all the way to the chat with a broken route to the model, and only then did you start hunting for the reason. Local running got a managed llama-server instead of node-llama-cpp, and Gemma 4 as the default model.
The interface is rewritten around the conversation. The Overview landing page is gone, and threads live in the sidebar. A request to approve an action arrives in the conversation that triggered it, and a thirty-day history of decisions is kept. Next to the chat you can open git diffs, a browser, a terminal, and a file editor. That editor can only modify existing files, not create or delete them, and the changes panel is read-only. The /btw command opens a side conversation, so you can ask a silly question without cluttering the main thread.
The interface loads faster. From 140 JavaScript requests down to 45, startup from 1.6 seconds to 575 milliseconds. But that is a simulated test with a faked gateway and an artificial latency of 50 milliseconds, so a lab number. On your server it may not look like that.
Sessions and transcripts moved to SQLite. Careful when rolling back to an older version: sessions created after the migration won’t show up in the old version. Make a backup before updating, and one you’ve verified.
Multiplayer, the headline feature
Multiple people can sit in one thread at the same time. A session has an immutable founder, an assignable owner, and a list of people who have written into it. The owner or an admin sets for each person whether they can only read, propose changes, work in a draft, or participate fully. The interface shows who is present and who is typing right now, and each message shows its author. You start a new session as a draft, so it doesn’t appear in colleagues’ panels until you publish it. Admins will see it anyway, though, with a grayed-out marker.
The OpenClaw team built this for themselves, and according to the announcement, OpenClaw itself is now being built on it. Inviting a colleague into work in progress, or handing it over together with the memory of what the agent already knows, makes sense, and it wasn’t possible in the single-user version.
And now read it in the documentation
The security documentation says: one gateway, one trust boundary. Everyone who can write to the agent shares its tool permissions, its credentials, and its files. The permissions from the previous section are a coordination aid for a team that already trusts each other. It is not tenant isolation, and the documentation says so openly. The release notes add two details that I would want to know in advance: revoked access may look valid for a while until the interface refreshes, and author attribution for individual steps is only best effort.
Add to that the default security posture. Command execution on the host runs without approval, and the sandbox is optional and has to be turned on manually. For one person on their own laptop that is reasonable. The moment you let a second person into the thread, it is a shared shell. What an agent can do with that kind of power in someone else’s system, we showed in the case of the deleted reservation.
And one trap that catches everyone who exposes an agent on Telegram. By default, direct messages from all people fall into one main session. That is, into one shared context. You want dmScope: "per-channel-peer", otherwise your users will be reading each other.
My take: multiplayer is a good feature that’s poorly communicated. If the announcement had contained a single sentence saying that a shared gateway is one trust boundary, I’d be satisfied. As it is, readers have to find that in the documentation themselves, and most won’t.
Where Hermes still leads
Hermes Agent from Nous Research, released in February, bets on depth. It can write skills for itself from work it has done and gradually improve them, and its memory is a fact database in SQLite with full-text search. It grows as you use it.
OpenClaw has had Skill Workshop since June. The agent proposes a skill as PROPOSAL.md, and you approve it, reject it, or send it back for rework. Version 2.0 didn’t change anything there. OpenClaw’s memory is markdown files indexed into SQLite, in other words, search over notes.
I’ll admit honestly, though, that I consider the approval step the better idea. A bad answer is one bad answer, while a bad skill spoils every subsequent run. The cost is clear: the agent doesn’t improve unless you’re sitting there.
Incidentally, the most telling item in the whole release is in the section on migrations. OpenClaw 2.0 can import an existing Hermes installation. As recently as this spring, it worked the other way around.
Who should turn it on
A family or a single team in one company: go ahead, it’s a supported deployment and it works. Clients, freelancers, and anyone you don’t want to give the keys to the whole machine: a separate gateway, ideally a separate operating system user or another server altogether. For a self-employed person who leaves an agent with access to their accounting and invoices, this is the difference between comfort and disaster.
And whoever wants an agent that improves itself should stay with Hermes. Version 2.0 doesn’t compete with it there.
Sources: the OpenClaw 2.0 announcement, Accidentally (Aug 30, 2026), the 2026.8.1 release notes, the gateway security documentation, and the documentation on multi-user mode and on memory. At the time of writing, the release notes are still marked as a draft in editorial review, so the wording may change. Data valid as of Aug 31, 2026.