28069db63a
* feat(macos): render system chat option cards * fix(macos): localize question card controls * test(macos): stabilize main actor attachment tests
6.6 KiB
summary, read_when, title, sidebarTitle
| summary | read_when | title | sidebarTitle | ||
|---|---|---|---|---|---|
| First-run setup flow for OpenClaw (macOS app) |
|
Onboarding (macOS app) | Onboarding: macOS App |
The macOS app's first-run flow: pick where the Gateway runs, connect a verified AI backend, grant permissions, and hand off to the agent's own bootstrap ritual. For CLI onboarding and a comparison of both paths, see Onboarding Overview.
Security trust model:
- By default, OpenClaw is a personal agent: one trusted operator boundary.
- Shared/multi-user setups need lock-down: split trust boundaries, keep tool access minimal, and follow Security.
- Local onboarding defaults new configs to
tools.profile: "coding"so fresh setups keep filesystem/runtime tools without the unrestrictedfullprofile. - If hooks/webhooks or other untrusted content feeds are enabled, use a strong modern model tier and keep strict tool policy/sandboxing.
Where does the Gateway run?
- This Mac (Local only): onboarding configures auth and writes credentials locally.
- Remote (over SSH/Tailnet): onboarding does not configure local auth;
credentials must already exist on the gateway host. The remote gateway token
field stores the token the macOS app uses to connect to that Gateway;
existing
gateway.remote.tokenSecretRef values are preserved until you replace them. - Configure later: skip setup and leave the app unconfigured.
- Gateway auth mode defaults to
tokeneven for loopback binds, so local WS clients must authenticate. - Setting
gateway.auth.mode: "none"lets any local process connect; use that only on fully trusted machines. - Use a token for multi-machine access or non-loopback binds.
Once the Gateway is ready, onboarding looks for AI access you already have:
a Claude Code or Codex login, OPENAI_API_KEY / ANTHROPIC_API_KEY, or a
tool-capable model already installed in a reachable Ollama or LM Studio server.
Detection runs on the Gateway host, including when the macOS app connects to a
Linux Gateway. The best option is tested with a real completion and only saved
after it answers; when a test fails the app automatically tries the next option
and shows why the previous one failed. If several options are found you can
switch between them before continuing. Automatic local discovery never pulls
or downloads a model.
To use a Claude subscription when the Gateway host has no Claude CLI login, run
claude setup-token on any machine with Claude Code installed, then paste the
printed token as Anthropic setup-token under Connect with an API key or
token.
Installed Gemini CLI, Antigravity, Pi, and OpenCode CLIs are shown for context when they cannot be selected as the reusable guided-setup inference route. Gemini and Antigravity cannot enforce the tool-free inference probe. Pi and OpenCode are whole-agent harnesses rather than setup inference routes; their session integrations require separate runtime and plugin setup.
You can also sign in through the provider's own OAuth or device-pairing flow. The built-in choices include OpenAI/ChatGPT, OpenRouter, GitHub Copilot, Google Gemini CLI, xAI, MiniMax Global and CN, and Chutes. The list comes from the Gateway's active text-inference provider plugins rather than a fixed app list, so another provider can opt in without adding provider-specific macOS code.
The manual key/token picker uses the same provider registry. In every route,
the provider supplies its starter model and configuration; OpenClaw verifies
the credential with the same live test before storing its auth profile. Next
remains locked until one backend has passed, so the first agent chat cannot
start without working inference. After that live check passes, OpenClaw becomes
available to help configure the remaining workspace, Gateway, channels, and
other optional features. When OpenClaw offers a short list of choices, the app
shows native option cards; choosing one sends the selection, and Skip for
now always leaves the choice optional. OpenClaw is also available later under
Settings → OpenClaw.
For a local Gateway, onboarding checks the Mac for memories from supported AI
tools: Claude Code auto-memory, Codex consolidated memories, and Hermes memory
files. When any are found, this page lists each source with its memory count
and lets you import the selected sources into the agent workspace under
memory/imports/ for indexed recall. Already-imported files are skipped, and
the page never appears when there is nothing to import. Skipping is safe; the
dashboard's Memory import page offers the same import later with per-file
control.
Onboarding requests TCC permissions for: Automation (AppleScript), Notifications, Accessibility, Screen Recording, Microphone, Speech Recognition, Camera, and Location.
After inference passes, OpenClaw owns the remaining optional setup and can hand you off to the normal agent chat. Finishing the permission walkthrough opens that same chat; the app does not create a workspace or launch a separate agent setup conversation before OpenClaw. See [Bootstrapping](/start/bootstrapping) for what happens on the gateway host during the agent's first real turn.