← Back to OpenClaw News A layered paper-cut illustration of a compact gateway arch inside a small home server cabinet, a blank signpost waiting for a choice of road, and paper lanterns rising from a lab bench into open sky
November 22, 2026Release WatchSecurityEnterpriseSkills

OpenClaw Slims Its Gateway for Small Machines, Stops Defaulting Chats to OpenAI and Graduates Code Mode From Labs

No new release on Saturday, but main got a lean Gateway profile that shares one worker host and holds growth to a memory budget, stopped quietly sending model-less chats to OpenAI, and graduated Code Mode, Tool Search and Templates. Owners can now unlink a claw from OpenClaw ID for good, and Git installs update through prepared releases.

No new OpenClaw release shipped on Saturday. When we checked at about 6:30 AM Sunday, npm still listed 2026.11.6 as latest, 2026.11.10-beta.1 as beta and 2026.10.33 as extended-stable, and the GitHub releases list had nothing newer than Thursday's tags. Main was busy, though, with 244 pull requests merged between 4 AM Saturday and our check. Three changes stand out. Small machines get a new lean resource profile that shares one host across the Gateway's core workers and holds worker growth to a memory budget. A fresh install with no model chosen will no longer send chats to OpenAI behind your back. And Code Mode, Tool Search and Templates graduated out of Labs. We also cover a permanent way to unlink a claw from OpenClaw ID, a rebuilt update path for Git installs, a security tip on model defaults and launcher ownership, and a skill card for keeping an eye on blogs. All times are Pacific, and none of these merges is in a release yet.

A Lean Gateway for 1–2 GiB Machines

The groundwork landed Friday at 3:14 PM, when pull request #202422, which shows Gateway resource profiles and live worker memory, added a gateway.resources.profile setting with three values: auto (the default), standard and lean. Auto picks lean when effective memory is "strictly below 8 GiB," and openclaw status now reports the chosen profile, the reason and a live census of workers. That PR said plainly that it did not change memory use on its own.

The two changes that do merged overnight. At 11:21 PM Saturday, pull request #202495, which runs lean core workers in one shared host to cut small-machine memory, tackled a simple problem: "Each worker loads its own copy of largely the same code, so worker memory grows past what a 1–2 GiB host can hold." In lean, the SQLite, transcript, state-read and catalog workers now run as endpoints inside one shared host worker. They load common code once but keep their own ports, queues and state. The trade-off is stated up front: if a hosted task fails, is cancelled or times out, the whole shared host is replaced, and in-flight work on every hosted endpoint fails with an explicit "Shared worker host failed" error rather than hanging. In the PR's benchmark on 1.5 GiB containers, the personal-assistant workload finished all three lean rounds with no out-of-memory kill, while the standard layout finished one of three. Lean's busy-phase Gateway memory was lower in all 10 paired rounds.

At 12:36 AM Sunday, pull request #202507, which keeps lean Gateway worker growth inside a memory budget, added the second half. In lean, an extra worker starts only if it fits the budget; otherwise the work waits for a worker that is already running. The default is 85% of a cgroup v2 container limit, or 40% of host RAM when there is no container limit, and a new gateway.resources.memoryBudgetMiB setting overrides both. The PR is careful to call this "an admission rule, not an OOM guarantee," and its own test showed floor workers alone could reach a 1.5 GiB limit. The OpenClaw Gateway configuration reference already documents both settings, says changing the budget requires a Gateway restart, and repeats that "Neither profile is a memory cap or OOM guarantee."

SEN-X Take

This is the most practical change of the weekend for anyone running OpenClaw on a small VPS, a home server or a 4 GiB container. Auto already picks lean below 8 GiB, so most small hosts will get it without touching config once a release carries it. Do not treat it as permission to shrink your box, though: the maintainers' own numbers show 1.5 GiB is still tight. Check openclaw status after upgrading to see which profile you landed on.

No Model Chosen Now Means No Model Used

At 12:11 AM Sunday, pull request #203271, which stops sending chats to OpenAI when no model is configured, merged. It describes a quiet default: "On a fresh install with no model chosen, OpenClaw silently behaves as if openai/gpt-6-astra were configured," and chat turns "are sent to OpenAI even though the user never picked OpenAI or signed in to it." The Models page showed an error about loading OpenAI models, and a utility model was derived from a primary model that nobody had chosen.

Now OpenClaw no longer picks a model for you. With none configured, a chat turn gets a setup message telling you to choose one in Settings → Models, send /model <provider/model> in chat, or run openclaw models set <provider/model>, and this happens before any model call. Commands that need no model, such as /help, /status and /new, still work. Existing installs are protected: when updating from 2026.11.6 or earlier, openclaw doctor --fix writes the model the install was actually running as agents.defaults.model.primary, and it writes the OpenAI model only when OpenAI credentials are present. For plugin authors, resolveDefaultModelForAgent now returns undefined when no model is set, and DEFAULT_PROVIDER is deprecated. The OpenClaw agent model configuration docs now say that if no provider lists a model, "the agent has no model and turns ask you to choose one."

SEN-X Take

A self-hosted agent should never pick a cloud provider on your behalf, so this is the right fix, and an overdue one. For most teams the practical risk was small, because a turn without OpenAI credentials would simply fail. But anyone who had an OpenAI key in the environment for some other reason could have been sending conversations to a provider they never chose. Run openclaw models status today and make sure the primary model is the one you mean.

Code Mode, Tool Search and Templates Leave Labs

At 8:16 AM Saturday, pull request #203096, which moves Code Mode, Tool Search and Templates out of Labs, merged. The PR notes that Code Mode and Tool Search "already have core defaults," but their controls were still presented as experimental, and Templates already had a core catalog, creator and installer that a missing Labs opt-in kept hidden. Code Mode and Tool Search now live under Settings → Agents → Agent defaults → Tools, and Templates is available through Customize → Templates without writing anything new into saved configuration.

Nothing about the defaults changes. According to the OpenClaw experimental features page, an absent tools.codeMode still uses "auto" "for catalog-preferred models only," and the switch "selects Auto or Off without forcing unsupported models on." Tool Search "remains enabled by default when tools.toolSearch is unset." For Templates, an existing explicit false stays off, and installation "still requires its normal authority, review, and consent." The PR adds that graduation "does not grant installation or publication authority" or change risk acknowledgements. Its upgrade test started from 2026.11.6 and checked that a missing Templates setting changed from denied to allowed while an explicit false stayed denied.

SEN-X Take

Graduation is mostly a signal: the maintainers now treat these three features as part of the product, not experiments. The one behavior change to plan for is Templates showing up for everyone who never set the old Labs key. If you run OpenClaw for a team and do not want people browsing and adding Templates yet, set the advanced opt-out, gateway.controlUi.experimental.templates, to false before you upgrade, not after.

Unlink a Claw From OpenClaw ID for Good

Until now, an owner had no supported way to take a claw off OpenClaw ID. Pull request #201750, which adds the OpenClaw ID protocol for unlinking a claw from its own computer, merged at 11:30 AM Saturday and noted that the only option had been "an operator script with a Cloudflare token." The user-facing half followed at 10:13 PM: pull request #201779, which lets owners unlink an OpenClaw ID claw from its own computer, adds openclaw id unlink and an Unlink from OpenClaw ID item in the account menu.

Before anything happens, the owner sees how many people and devices will lose access, that the address stops working at once, and that the name "can never be used again, by any account." Every owner device, member, node and invite is removed, remote sessions close and the tunnel stops, but conversations stay and the claw keeps working locally. The action works only on the computer that runs the claw; over the tunnel the Gateway refuses it. If OpenClaw's Discovery service cannot be reached, the claw is still unlinked locally and the signed unlink is sent later, with openclaw id status showing "Unlink pending." The OpenClaw ID CLI reference already describes the command, including --yes to skip the confirmation.

Git Installs Update Through Prepared Releases

OpenClaw's update work continued. At 1:23 AM Saturday, pull request #200578, which runs every Git install from releases and deletes the in-place Git runner, removed the old updater that rewrote the checkout a running Gateway uses. The PR calls in-place replacement "RC1 of the update RCA," the first root cause listed in the team's root-cause analysis of updates. Windows checkouts, package-to-Git switches and checkouts with no Gateway service now use the same release path as other Git installs. At 11:47 PM, pull request #203194, which updates Git installs through prepared releases and a resident launcher, spelled out the flow: build the next release beside the running one, test it offline against current config and state, publish it as a trial, then let the launcher commit it once the new Gateway is ready or roll back to the previous release.

A smaller fix addresses an expensive failure loop. Pull request #203236, which stops automatic updates from retrying the same failure every hour, merged at 11:16 AM Saturday. On one affected install, it says, automatic updates had failed the same way since November 3, building up about 330 failed runs, each costing about 10 minutes of validation and a roughly 1.9 GiB state snapshot. Identical failures now back off to 1, 2, 4, 8 and 16 hours, then once a day.

SEN-X Take

Updates have been OpenClaw's weakest link for months, and these changes go after the root cause instead of adding more recovery steps. For Git-based installs, the first update under the new design converts the existing install to the release layout, so run that update while someone is watching. If your logs show the same automatic update failing every hour, a manual openclaw update resets the backoff.

Security Tip: Name Your Model, and Check Who Owns Your Launcher

Skill Card: Blogwatcher by Peter Steinberger

Browsing skills by topic is about to get easier: at 3:31 PM Saturday, pull request #202756, which lets the Control UI browse and search Skills by ClawHub category, added category sections with View all and search scoped to a category. One skill in ClawHub's Research category is a simple, useful pick. The Blogwatcher skill page on ClawHub, published by @steipete, says it will "Monitor blogs and RSS/Atom feeds for updates using the blogwatcher CLI." When we checked, it showed version v1.0.0, about 45,300 downloads, 75 bookmarks, a security audit result of "Pass," an MIT-0 license and a last update six months ago.

  • What it does. The SKILL.md teaches an agent a handful of commands: add a blog by name and URL, list tracked blogs, scan for updates, list articles, and mark one or all of them read. Pair it with an automation and an agent can check your industry blogs every morning and summarize only what is new.
  • Install the right one. ClawHub's API reports four skills with the slug blogwatcher, so a bare name is ambiguous. Use the full reference from the listing: openclaw skills install @steipete/blogwatcher.
  • Our caveats. The skill relies on a separate tool. Its setup step runs go install github.com/Hyaxia/blogwatcher/cmd/blogwatcher@latest, which pulls the latest code from a third-party GitHub repository rather than a pinned version. Review that repository and consider pinning a tag before letting an agent install it. We read the listing; we did not install it.

Also on Main: Harness Choice, Automations and Trusted Proxies

A few smaller merges are worth a line each. At 5:44 PM Saturday, pull request #203358, which lets Settings choose a harness for primary and fallback models, lets you pick OpenClaw or Codex in either dropdown when the catalog offers both; before, you could choose a harness only in chat. At 2:34 AM Sunday, pull request #203106, which keeps failing automations running and lets triggers wait for their commands, fixed a chat-created GitHub watcher that "went dark after its first scheduled run and nobody was told": a slow gh search returned a still-running handle, the scheduler treated the script error as permanent and turned the job off without notice.

Earlier, at 8:56 AM Saturday, pull request #199699, which connects approved trusted-proxy CLI and TUI devices within their grant, added an opt-in gateway.auth.trustedProxy.narrowToDeviceGrant setting, off by default, so a paired CLI or TUI guest connects with only the scopes it was approved for instead of filing a scope-upgrade request. And at 9:14 PM, pull request #203408, which rejects cross-conversation automation workspaces before saving, makes an unsupported automation target fail when it is saved, rather than on every scheduled run.

What We Left Out

The official OpenClaw account posted on X on Saturday morning that Omar Shahine's Autopilot, "Sebastian," drafted 40 Teams replies and sorted them by priority while he was out; we covered Autopilot on Wednesday, so we did not repeat it. We did not cover user summaries of 2026.11.6, which we reported on Friday. We also skipped a Claw Studio update, an Acurast deployment post, a Penfield memory plugin release and several "best agent" comparison threads, since we could not confirm them against a primary source in time, and MyClaw 3.0, which is not an OpenClaw product. Pull request #203389, which reorders the curated model shortlist, is a catalog ranking change and will matter only after a hosted catalog refresh.

What to Watch Next

Watch for the next beta or stable tag that carries the lean profile, the end of the silent OpenAI default and the Labs graduations. Also watch whether OpenClaw publishes small-machine guidance alongside the lean profile, and whether package installs move to the same prepared-release update path that Git installs now use.

Running OpenClaw on a Small Server?

SEN-X helps teams size Gateway hosts, pin the models their agents use, and plan OpenClaw upgrades so updates and rollbacks never catch them off guard.

Contact SEN-X →