Cloud providers

Cloud AI is the ai node with a cloud provider. There is no separate cloud node type. Cloud egress is a deliberate carve-out from Flow’s zero-egress posture. It defaults to off, and users opt in via Settings.

Posture

	`ai` node, `local` provider	`ai` node, cloud provider
Network egress	None	Yes - to the provider’s API
Credentials	None required	Provider API key (OS keyring, env fallback)
PII sanitization	Mandatory	Mandatory (same sanitizer)
Default policy	Always on	Off; opt-in via Settings
Persisted output	Full assistant text	Metadata + 200-char preview by default
Visual badge	None	A cloud egress indicator on the node

Supported providers

Provider	Env var fallback	Example default models
Claude (Anthropic)	`ANTHROPIC_API_KEY`	Opus, Sonnet, Haiku families
OpenAI	`OPENAI_API_KEY`	GPT-4-class models
Gemini (Google)	`GOOGLE_API_KEY`	Gemini 2.x / 1.5 families
NVIDIA	`NVIDIA_API_KEY`	Nemotron and hosted open models

Adding a new provider is a contained, code-level change. See Extension APIs.

Capabilities, at parity

The same five capabilities wire through to each provider in its native format. Reasoning, vision, and tool use work on the major providers. Embeddings work where the provider exposes an embeddings API, which Claude does not. Classification works everywhere. Structured output rides as native JSON-schema enforcement on OpenAI-compatible providers, and is prompt-constrained elsewhere. Agentic mode works with any provider. Generation uses the node’s provider and model, and review, monitor, and fix are the shared lifecycle.

Privacy and audit

Metadata-only by default. A cloud node’s persisted output contains provider, model, finish reason, token counts, latency, and a 200-character preview. This supports cost and latency review without keeping transcripts on disk. Setting auditContent: true on a node persists the full prompt and response for debugging or training-data collection.

What flows out: the provider receives the sanitized prompt. Datasets, hostnames, credentials, and IPs are replaced with typed placeholders before the HTTP call. Each provider has its own data-retention policy, and Flow does not negotiate or alter those terms.

Credentials

API keys resolve at execution time through keyring, then env var:

Keyring (recommended) is stored in the OS-native credential store. Set it from Settings → Providers. The key is never echoed back, and the row shows a keyring-OK indicator.
Environment variables (fallback) are the matching env var per provider. The Settings drawer marks env-only providers with an amber hint to migrate.

If neither path produces a key when a cloud node runs, the node fails with a clear message naming the provider and the env var. The runtime never persists keys to disk outside the OS keyring.

Settings

{
  "allow_cloud_ai": false,
  "providers_enabled": {
    "claude": true,
    "openai": true,
    "gemini": true,
    "nvidia": true
  }
}

The global toggle plus per-provider toggles live in the Settings drawer.

AI overview is the local-first counterpart.
Isolation boundaries show where the carve-out sits in the security model.
flow-execution is the internal dispatch path an ai node runs through.