Provider profiles and API keys
Configure named provider profiles, secure API credentials, local or remote execution surfaces, model slots, and per-vendor defaults.
Profiles are execution identities
Open Settings → Providers. A provider profile names one concrete way to execute a vendor or runtime: an API key, Application Default Credentials, a loopback service, a GCP IAP tunnel, or a remote GPU controller.
You can keep several named profiles for one vendor, but one is the effective default. A saved Flow selects a catalog Variant and a compatible profile at execution; it does not store the raw secret in the graph.
Create and verify a profile
- Select the vendor or service.
- Choose + Profile and give the profile a descriptive name.
- Choose the advertised authentication or execution mode.
- Enter only the fields declared by that mode.
- Choose Save settings.
- Choose Make default if this should be the vendor’s default execution identity.
- Choose Re-check and read the resulting status before selecting a paid Variant.
Changing a mode changes the contract. For example, Google API key and Google Application Default Credentials use different execution paths; a working AI Studio key does not prove Vertex AI readiness.
Where secrets are stored
API keys and controller credentials are stored in the server machine’s OS credential store. The browser receives only masked status and a tail, never the full stored value. A key saved in a profile takes precedence over the corresponding environment variable.
Enter credentials only in the declared Settings field or the approved setup terminal. Never paste a Hugging Face token, Google credential, Runpod key, or controller secret into Canvas, a prompt, or an agent chat.
Profiles imported from a Kumiho Desktop-owned installation can appear as Installed profile and be read-only here. They can be selected, but their credentials and connection fields must be edited in the owning installation. A local profile can also be temporarily locked while an installed GCP controller is the effective default.
Current provider families
The server-advertised catalog currently includes profiles for families such as:
| Family | Typical surface |
|---|---|
| Google · Gemini / Veo | API key or Vertex AI through ADC |
| OpenAI · GPT Image | Provider API key |
| BytePlus Ark · Seedance | Provider API key |
| Alibaba Cloud Model Studio · Wan | Provider API key |
| Kling AI | Provider API key |
| Black Forest Labs · FLUX | Provider API key |
| Topaz Labs | Provider API key |
| ElevenLabs | Speech, sound-effects, and music API key |
| MiniMax API | H3 video and supported MiniMax API capabilities |
| ComfyUI service | Local endpoint, GCP IAP, managed remote controller, or Runpod |
This is an orientation list, not a permanent compatibility promise. Settings is rendered from the running server’s capability catalog, and the selected Canvas Variant declares the profile modes it accepts.
ComfyUI profiles and model slots
A ComfyUI profile can point at the bundled local inference service, a local or trusted-network endpoint, a private GCP VM forwarded through IAP, an on-demand controller, or a Runpod controller. The same saved ComfyUI Variant can move between compatible execution surfaces without embedding credentials in the Flow.
One ComfyUI instance can host several model families. When the service is reachable, Settings asks that instance which files are legal for each model slot and presents pickers instead of trusting an arbitrary filename. Connect each model family’s files separately and resolve every missing slot before generation.
Use ComfyUI and GPU setup before opening a bundled installer. Raw ComfyUI has no built-in authentication; keep direct :8188 access on loopback or a trusted network.
Reading status correctly
- configured means the selected mode’s required credential or settings surface is present. For ADC this can be an ambient credential; for an endpoint it can be a saved URL. It is not proof of reachability or a successful model run.
- pending means the declared credential surface has not finished checking; choose Re-check.
- An unreachable local service can still leave the saved profile intact for later recovery.
- A model or region can be unavailable even when authentication succeeds.
- A paid provider call still requires the normal validation and spend gate.
For Google Cloud project credentials, continue with Google Cloud ADC and Vertex AI. For the meaning of Variant, Run, and Take, see Production records and glossary.