How it works
Watchflows calls the claude binary on your Mac instead of an HTTP API. You sign in to Claude Code once with your Claude account, and the app uses that session. Everything an AI node can do works here: plain prompts, JSON-schema structured output, images, and the Agent node's full tool loop.
Set it up
- Install Claude Code from
claude.com/claude-code. - Run
claudein Terminal and log in. - Add the provider: Settings ▸ AI Providers ▸ Add Provider ▸ Claude (Subscription).
The editor shows no key or URL field. In their place sit three live checks — Claude Code installed (green shows the version; red offers an install link), Signed in (red shows the copyable claude command), and Ready. Recheck re-runs all three at once instead of waiting for the next probe.
Models
The picker lists aliases, not pinned versions, so each one follows Anthropic's current model of that family without a Watchflows update.
| Alias | Resolves to |
|---|---|
sonnet | The current Sonnet model — the balanced default |
opus | The current Opus model — the most capable |
haiku | The current Haiku model — the fastest and lightest |
fable | The current Fable model |
Good to know
This is not a local provider. The claude binary is on your Mac, but it ships every prompt to Anthropic under your account. For nothing-leaves-the-machine flows, use Ollama, LM Studio, or Apple Intelligence.
Watchflows adds no charge. Calls draw on your Claude plan's usage limits, the same pool as your other Claude Code work. AI cost tracking reports this provider as free, because there is no per-token bill to meter.
Budget for turns, not calls. Each call takes seconds, at most two run at once, and every turn of an Agent loop is its own call against your plan. Temperature and max-tokens are ignored — Claude Code manages generation itself.
Watchflows needs CLI version 2.1.199 or later. An older binary fails the installed check with the version it found. Update Claude Code, then Recheck.