← Back to OpenClaw News An editorial illustration of a glowing browser window with a hand-shaped cursor stepping in while an automated cursor waits paused at the edge, beside a small software package running its own checklist on a test bench
November 14, 2026BrowserRelease WatchSecuritySkills

OpenClaw Lets You Take Control of the Agent's Browser as New Releases Rehearse Their Own Updates and Beta 3 Nears

A new Take control button lets you work in the agent's browser without it acting over you, and agents can now pause for a required sign-in instead of quitting. New releases will rehearse their own updates, 2026.11.5-beta.3 moves through qualification, OpenClaw ID reaches the Mac app, and a public benchmark prices OpenClaw against other harnesses.

Friday was a busy day on OpenClaw's main branch, though no new release shipped. The biggest change for people who let agents drive a browser is a new Take control button. It lets you work in the agent's browser without the agent clicking over you. A companion change lets an agent pause for a required manual step, such as a sign-in, instead of giving up on the task. Updates also changed hands: new releases will now run their own pre-update rehearsal, so a fix to those checks helps the update that ships it. Release tooling shows 2026.11.5-beta.3 moving through qualification. OpenClaw ID took a step toward everyday use on the Mac, and an open-source benchmark from Garry Tan put a price on how OpenClaw compares with other agent harnesses.

Take Control: You Can Now Pause the Agent's Browser and Work in It Yourself

At 11:41 AM Pacific on Friday, main merged pull request #197928, which lets users take over the agent's browser from the Browser panel. The PR describes a familiar problem: when you step into the Browser panel to solve a CAPTCHA, finish a login or fix a form, "the agent can click, type or navigate over you in the same browser, and it doesn't know you are there." Until now, the only way to stop it was to stop the whole run.

There are two parts. First, when you click or type in the Browser panel, the agent's browser results over the next minute note that you just used the browser, so it can check the page before acting. Second, Take control in the panel toolbar pauses browser actions from everyone else on that browser: the agent's browser tool, the openclaw browser CLI and other Control UI windows. The agent gets a normal tool result saying you have taken over, and its turn continues. It can still take snapshots and screenshots, and an action that had already started may finish. The updated OpenClaw browser profiles documentation describes the same behavior.

Control ends when you click Release control, when the Control UI window closes or loses its connection, or after five minutes without input from you. The lock belongs to the browser, not the profile name, so several profiles that reach the same browser through the same cdpUrl share one lock. The feature works only for browsers on the Gateway host. The button is hidden for node-routed browsers, and there is no new config option.

A second change, pull request #196951, which supports required manual-step handoffs in chat, merged at 5:37 PM. It says agents could end an authorized task when "a required human step blocks them," and that finishing a sign-in could leave the waiting task with its old, unauthenticated tool list. The existing ask_user tool now accepts a requiredAction. The chat question panel explains the blocker and the manual step, can open a verified link, and keeps the original task waiting. The OpenClaw ask_user documentation says Ready to recheck makes the agent re-observe the original target and refresh its tool list. It also says the answer "is not proof of authentication, permission, or successful completion, and it grants no additional authority." Stop this task, Skip and expiry leave the requirement unresolved and keep dependent work blocked.

The bundled browser skill changed to match. Pull request #198457, which removes blanket manual-step stop guidance (12:56 PM), deletes one line that told agents to "stop and tell the user exactly what is needed" whenever a page needed a login, permission, CAPTCHA, 2FA or another manual step. The PR says the line conflicted with the browser login docs, which allow owner-authorized credential fills. All other login, permission and Google Meet guidance is unchanged.

SEN-X Take

This trio fixes one of the most awkward parts of browser agents: the moment a human has to step in. Take control gives you a real lock instead of a race with the agent, and the handoff keeps the task alive without granting extra power. Note the order of operations. With the blanket stop line gone, agents may attempt more owner-authorized logins on their own, so review what credentials your browser profiles can reach.

New Releases Will Rehearse Their Own Updates

Our Friday report covered updates that roll back after a failed activation. The newest change moves earlier in the process. Pull request #197922, in which new releases run their own pre-update rehearsal, merged at 6:09 PM. Its problem statement is short: "A fix to the pre-update rehearsal cannot help the update that needs it." The installed updater runs its own copy of the state snapshot, Doctor rehearsal and test Gateway, so a corrected check has only reached users one update later.

Now, releases that declare openclaw.updateRehearsalProtocol: 1 in their package rehearse themselves. After staging, the installed updater runs the candidate's hidden openclaw update rehearse command in place of its own rehearsal. The openclaw update CLI reference now has a "Candidate-owned rehearsal" section. It says the candidate "receives no update or service authority" and touches copied state only. A pass grants nothing: the installed updater still owns drain, activation, verification and rollback. If the candidate's rehearsal crashes, goes silent or returns a result the updater cannot read, the updater names the reason and runs its own rehearsal, so the worst case is that rehearsal runs twice. This only starts paying off after an installation is already running a release with the change. Updaters from earlier releases ignore the new marker.

A related fix matters for long-term-support users. Pull request #198429 fixes 2026.9.x updates that report failure after installing the next release (2:37 PM). Updating from published 2026.9.x releases, including extended-stable 2026.9.37, to current main failed with post-update-failed and "Shared state migration did not finish," even though the update had installed and the Gateway was healthy. The PR says users on 9.36 and 9.37 can now update to the next release without a false failure or a manual openclaw update repair.

SEN-X Take

Letting the new version check itself sounds risky, but the design keeps authority with the installed updater. The candidate can only refuse, never approve its own activation. For teams on extended-stable 2026.9.37, the false-failure fix is the more urgent item. Wait for it to ship in a tagged release before you plan the jump, and rehearse on a staging copy first.

2026.11.5 Beta 3 Moves Through Qualification

No release appeared overnight. When we checked at about 6:30 AM Pacific on Saturday, the npm latest tag was still 2026.11.3, extended-stable was 2026.9.37 and beta was 2026.11.5-beta.2. The newest tag on GitHub was still v2026.11.5-beta.2, and 2026.11.4, which entered release qualification late Thursday, is still untagged. Release tooling points to the next beta. Pull request #198442, which authorizes a beta.3 Kova performance waiver (12:42 PM Friday), says the release owner approved waiving Kova performance failures for 2026.11.5-beta.3, keeping the failed CPU, latency and memory evidence on record. At 5:48 AM Saturday, pull request #198870, which honors accepted beta.3 validation failures, added the owner's acceptance of five failing jobs, including UI checks, Android screenshots and a Docker update test. The PR says these exceptions apply only to beta.3 and those five jobs, and calls it "not an unrestricted force-publish switch." Missing or cancelled jobs, unrelated failures and artifact integrity still block.

OpenClaw ID Comes to the Mac App and Lets Members Add Devices

Three OpenClaw ID changes landed Friday. Pull request #197630 lets you sign in with OpenClaw ID and approve a Mac as your device (2:34 PM). The Mac signs in, lists your claws and becomes an owner device of the one you choose, either by comparing six digits with a device that already has access or by entering the recovery kit. It then reaches chat and the dashboard "with its own key," with no VPN and no Gateway token. Moving an existing Mac to OpenClaw ID replaces its old connection and clears its saved Gateway token and password. The PR notes that this was not live-testable yet, because a staging service still needs a deployment.

Pull request #198575 lets members add and approve their own devices (9:02 PM). Before, an invited member could use only the browser that accepted the invite. Now members can approve more devices from one they already have. New devices inherit the member's role, owners see who approved each device on the People page and get a notice when a member adds one, and removing a device also removes the devices it approved. Members admitted before this change can still connect, but cannot approve devices until they are removed and re-invited. Earlier, pull request #198349 lets admins turn on remote access over a verified personal sign-in (10:21 AM), with the OpenClaw ID lab on, so setting up a remote claw no longer needs an SSH tunnel. Shared token or password connections and non-admins are refused.

SEN-X Take

OpenClaw ID is starting to look like normal account sign-in rather than network plumbing. The device-chain removal is the detail to like: pull one lost laptop and everything it approved goes with it. Still, remote access sits behind the OpenClaw ID lab, the member-device protocol is an unshipped draft that changed without a compatibility path, and the Mac sign-in was not live-testable yet. Pilot it on a test claw, and keep your recovery kit somewhere safe before migrating a working Mac.

A Public Benchmark Prices OpenClaw Against Other Harnesses

On November 9 we held back cost figures from Garry Tan's harness evaluation until we could read them in the repository. The g-harness-evals repository on GitHub, last updated on Friday, keeps the model and tasks the same and swaps only the harness. Its README says harnesses "differed more on cost than on how often they solved tasks." The g-harness-evals harness cost finding puts OpenClaw on Claude Opus 5.5 at $0.44 per solved task on 20 multi-user tasks, against $0.19 to $0.24 for Claude Code, Hermes Agent and QM. The README says this has no statistical test. It traces about 81% of OpenClaw's extra cost over Claude Code to cache writes, because OpenClaw re-sends a much bigger system prompt and bigger tool results. On GPT-6.1 Sol, the README says OpenClaw shares first place on cost with Codex on those multi-user tasks. The authors call these early "dev" results on public tasks for pinned harness versions, and the finding discloses that OpenAI is OpenClaw's key benefactor and that Peter Steinberger works for OpenAI.

Also on Main

Security Tip: Check Who Can Wake Your Discord Agent

At 12:13 PM Friday, main merged pull request #198331, which stops Discord presence greetings from waking the agent for members outside the guild allowlist. When presence greetings were on, a member excluded by the guild's or greeting channel's users/roles allowlist could still trigger an agent turn just by coming online. The PR says the presence path "was the one Discord ingress that skipped that check," so an excluded member could cause model inference and possibly an outbound greeting. It now uses the same member-access rule as messages, reactions and voice.

  • Know if you are affected. The OpenClaw Discord events documentation says presence events need intents: { presence: true }, the privileged Presence Intent and a presenceEvents block on the guild. If you never set those, this path was not active for you.
  • Treat the allowlist as the trigger list. The docs now say members "must also pass the guild or target-channel users/roles sender allowlist, like any other trigger." presenceEvents.users can only narrow that audience further. Guilds with no users or roles restriction see no change, so add one if strangers share the server.
  • Keep the throttles. The same page lists the eight-hour per-user cooldown, a 300-second quiet window after a new Discord Gateway session and a default limit of 8 events per 60 seconds per guild. Until the fix reaches a tagged release, these limits and a tight allowlist are your guardrails.

Skill Card: Browser Automation (Bundled)

With Take control and manual-step handoffs landing, it is a good time to read the browser-automation skill that ships with OpenClaw's browser plugin. Its description says to use it "when controlling web pages with the OpenClaw browser tool, especially multi-step flows, login checks, tab management, or recovery from stale refs/timeouts." Its frontmatter marks it as not user-invocable, so it guides the agent rather than acting as a command you run.

  • An operating loop. Check browser state first, prefer stable tab handles such as labels, read before you click with snapshot or text, act narrowly with refs from the latest snapshot, and report real blockers instead of looping.
  • Tab hygiene. It tells agents to list tabs and reuse a matching label or URL before opening new ones, and to close duplicates left by retries.
  • What changed Friday. Pull request #198457 removed the blanket "stop and tell the user" line for logins. The OpenClaw browser login guide still recommends manual login in the host browser's openclaw profile, and says an agent may use owner-supplied credentials for an owner-authorized task, keeping "credentials out of chat and replies."

What We Checked and Left Out

The Shenzhen Claw Hack and ClawCon posts on X repeat the December 4–6 dates we reported on Friday, so we did not cover them again. A maintainer reply on X about an upcoming change to cautious safety defaults gave no details, so we left it out. We could not find the TensorFold or veryfi-ocr skills promoted on X through ClawHub's public API, and we skipped secondhand reports about Telegram widget settings and resurfaced, months-old NVIDIA CVE threads. X posts described a full Boston meetup, but there was no primary write-up to cite.

What to Watch Next

First, whether 2026.11.5-beta.3 reaches npm, and whether 2026.11.4 finally ships as stable. Second, which release first carries candidate-owned rehearsal and the 2026.9.x false-failure fix, since extended-stable users need both. Third, whether Take control comes to node-routed browsers, and whether OpenClaw ID sign-in on the Mac leaves the lab.

Letting agents drive real browsers and accounts?

SEN-X helps teams set up browser profiles, human handoffs and Discord allowlists safely, and test OpenClaw updates on staging before they reach production agents.

Contact SEN-X →