ChatGPT (Subscription)

What It Is

ChatGPT (Subscription) is an AI provider that runs your AI nodes through Codex — OpenAI's command-line app for ChatGPT. Instead of pasting an API key, you sign in to Codex once with your ChatGPT account, and Watchflows uses that session.

Everything an AI node can do works here: plain prompts, structured JSON output with schema validation, image input (vision), and the AI Agent node's full tool-using loop. To Watchflows it's just another provider in the picker — the difference is who bills the usage (your ChatGPT plan) and how the call travels (through the local codex binary rather than an HTTP API).

Setup

1. 1 Install the Codex CLI — run npm install -g @openai/codex in Terminal if it isn't on your Mac already.
2. 2 Sign in — run codex login in Terminal and log in with your ChatGPT account.
3. 3 Add the provider — Settings → AI Providers → Add Provider → ChatGPT (Subscription).

There are no API key or URL fields. In their place the provider sheet shows a live status card with three checks:

• Codex CLI installed — green shows the installed version; red offers the copyable npm install -g @openai/codex command. Watchflows needs version 0.142.5 or later.
• Signed in — red shows the copyable codex login Terminal command that starts the sign-in flow.
• Ready — both checks pass and the provider can run.

Just installed or just signed in? Hit Recheck — it re-runs the checks immediately instead of waiting for the next automatic probe.

Billing & Limits

Calls consume your ChatGPT plan's usage limits — the same pool as your other Codex use (the IDE extension, the CLI, cloud tasks). When your plan's limits are hit, runs pause until the limits reset on OpenAI's schedule: the AI node fails with a clear plan limit reached message rather than looping or retrying on its own.

Because the limit is shared, an always-on automation that fires AI nodes frequently can draw down the same budget you rely on elsewhere. Keep an eye on high-frequency triggers, and consider a local provider for the busy, low-stakes steps.

Models

Three models are available. They share the same codex exec pipeline, so structured output and vision work on all of them — they differ in capability and speed. New nodes default to gpt-5.5.

Vision

Unlike Claude (Subscription), this provider accepts image input. Point an AI node's Image field at an upstream image — a screenshot, a fetched URL, a photo from the Photo Added trigger — and the model can see it and describe, classify, or extract from it.

When you add the provider, Watchflows sets a Default Image Model for you, so a node left on its defaults routes image-carrying payloads to vision automatically. You can also pick a specific model in the node's Model field. If you added the provider on an older build and see a no Default Image Model set message, open Settings → AI Providers, edit ChatGPT (Subscription), and choose a Default Image Model once.

Limitations

• Temperature and max-tokens are ignored — Codex manages its own generation settings.
• Requires the CLI installed and signed in — the provider is unavailable in the picker until both checks pass.
• Latency is seconds, not milliseconds — each call boots the Codex CLI, so expect a few seconds per call. Fine for automations; noticeable if a flow chains many AI nodes.
• AI Agent runs are metered per turn — every turn of an AI Agent's loop (each tool decision or reply) is its own Codex call, and each turn draws on your plan's usage. A long agent run spends many turns of plan usage where a single Prompt node spends one.
• Concurrent runs are queued — a couple of calls run at once; the rest wait their turn instead of stampeding your plan.

Troubleshooting

These are the exact messages the app shows, and what to do about each.

Privacy

Prompts and payload data go to OpenAI under your account, governed by your ChatGPT account's terms and data settings. Nothing routes through Watchflows servers — the app talks only to the Codex CLI on your Mac, which talks directly to OpenAI.

Unlike Ollama, LM Studio, or the Apple on-device model, this is not a local provider: your data does leave the machine. For fully offline or privacy-sensitive flows, use a local provider instead.