Skip to main content
Storage and downgrade warningThis release changes how sessions and transcripts are stored by moving them into SQLite. Before downgrading to an older file-backed release, use the current CLI to restore archived legacy transcript artifacts; sessions created after the migration will not appear in older releases.Create a verified backup before upgrading to protect broader OpenClaw state, and review downgrading across the session SQLite migration before rolling back.
OpenClaw now gives new Mac, Linux, and Windows installs a clearer path from download to a first useful conversation, while iPhone, iPad, and Android put pairing and permissions where people need them. Guided setup can reuse supported subscriptions, API keys, and local models already available, verifies the chosen model before saving it, and hands off to the web app or terminal when the connection is ready.
The supported install path now keeps the app or command available after setup. A Mac app opened from Downloads or a disk image can offer to move itself into Applications, where updates and launch at login work properly. On Linux and other Unix systems, the installer makes openclaw available in new terminal sessions without asking you to edit shell startup files by hand.Network installations that would expose OpenClaw without authentication are stopped before anything changes. Reinstalling also protects an existing working setup when preparation is cancelled or fails, and gives OpenClaw time to start before reporting whether it is reachable.
Guided setup now starts by looking for AI access you may already have. It can reuse verified Codex, ChatGPT, or Claude CLI sign-ins, accept an API key, run a supported provider’s own sign-in, or find qualifying Ollama and LM Studio models, then prove that the exact choice can answer before it keeps that model and credential. Access already working on the machine can become part of setup instead of another thing to configure.For OpenAI accounts, setup uses the models the signed-in account can actually access while preserving routes you configured yourself, and administrators can prepare supported local models with live progress. Finding or downloading a local model is only the start, so the supported local-model screens do not show Start chatting until that exact choice passes activation.
Successful setup now hands off directly to a first conversation. Graphical Mac, Linux, and Windows sessions can open the web app, while SSH and other headless setups provide an authenticated link with port-forward instructions and keep terminal chat available.From there, the setup conversation can finish supported skill and web-search configuration, then offer an external channel as an optional next step instead of another prerequisite for first use. The classic wizard can still continue without AI because its live model check remains optional.
On Mac, the main guide points directly to the app, and the local or remote Gateway you choose stays selected even if older startup or cleanup work finishes late. The app waits through the Local Network prompt and legitimate first-run data upgrades, authenticates the exact remote Gateway before moving on, and opens the dashboard only when the connection is ready.On Windows, the guide points to the latest signed x64 and Arm64 Hub installers. The PowerShell installer now recognizes a supported Node runtime correctly and can continue in the same session after Winget installs Node.js. Windows Hub updates independently, so its standalone stable build can be newer than the mirror included with an OpenClaw release.On Linux, desktop setup can repair or reinstall OpenClaw, connect to a local or remote Gateway directly or through SSH, verify eligible AI access already on the machine, and resume an interrupted activation. Direct certificate-pinned connections remain unavailable inside the desktop app.
Mobile setup now puts QR and setup-code pairing first because that is what most people were trying to find anyway, while discovered Gateways and manual host and credential entry remain available. Secure official pairing shows whether the phone has Full or Limited access, and an unencrypted connection to another machine is automatically kept Limited.On iPhone and iPad, a valid setup QR can now pair and open the main UI in one scan when the connection is already trusted or matches the code, while untrusted connections still stop for approval. OpenClaw explains pairing before asking for Local Network access, lets you decide on optional permissions one at a time, and gives eligible administrators a dedicated Settings conversation for Gateway setup and repair without exposing that privileged assistant in ordinary Chat.On Android, pairing stays usable in landscape, on narrow screens, and with larger system text. Public Gateways can use normal certificate checks while LAN and IP connections keep explicit pinning, and pairing again preserves saved location and notification choices instead of quietly changing consent; revoked permissions remain revoked, and Google Play builds still use foreground location only.
Imports from Claude, Codex, and Hermes now happen in a temporary staging area, where OpenClaw verifies or repairs the model route before making the new setup active. If the source, import plan, or destination changes along the way, promotion stops rather than replaying a half-finished import against different data.Rerunning onboarding keeps the workspace you already use unless you approve a move, and named agents keep the credentials they own through supported creation and configuration-only resets. OpenClaw validates the provider, Gateway, migration, and workspace choices before a reset can move existing data to Trash, but a valid confirmed reset is still destructive.
Non-interactive onboarding now rejects invalid provider, authentication, Gateway, workspace, and conflicting flow choices before it creates an agent or writes configuration. When --json is requested, failures return one machine-readable result instead of empty or mixed output, and a concurrent setup for the same profile fails quickly with the current holder when known instead of sitting there looking frozen.Cancelled or interrupted setup revokes stale work, while failed provider sign-in or local-model preparation can be retried without cancelling a newer attempt. Gateway health failures stay inside setup with the original diagnostic and recovery hints, but --json does not waive risk acknowledgement, and deliberately skipping service startup can still finish without a reachable Gateway, so automation must inspect the reported health instead of trusting exit status alone.