Connect a model provider
Give the assistant a brain: paste a cloud key, or point Zeph at a local runtime on your Mac.
A model provider is the thing that answers — the brain behind the assistant and behind any workflow step that runs Ask agent. Everything else Zeph does works without one; the moment you want it to think, you connect one.
Two shapes ship today: a cloud service you paste a key for (Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, Vercel AI Gateway), or a local runtime already on your Mac that Zeph discovers (Ollama, LM Studio, Apple Intelligence). You do it once, in the integrations library in the companion.
Paste a cloud key
Open the provider in the integrations library
Find the one you want — Anthropic, OpenAI, Gemini, Groq, OpenRouter, Vercel AI Gateway. Each shows a Get a key link straight to the vendor if you do not have one yet.
Paste the key
Zeph checks it with the vendor before it keeps it, and stores it in the macOS Keychain — never in a config file. If the key works, the row turns green and reads Connected. If the vendor rejects it, you get the reason and a Reconnect; if Zeph could not reach the vendor at all, you get Retry.
Pick a default model
Once connected, a default-model picker appears. Choose one, or leave it blank to use the provider's own default.
Point at a local runtime
Start Ollama or LM Studio on your Mac, then open its row in the library. Zeph already knows where each one usually listens, so the default endpoint is shown in the field — press Test and it reports Connected — N models available. Leave the endpoint blank to use that default, or type a custom one if you moved it. Then set a default from the models it discovered.
Apple Intelligence runs on the device itself. There is nothing to point at — its row simply reads Runs on this device, discovered automatically.
Which runtime holds the models
A local runtime brings its own models — you add and remove them in Ollama or LM Studio, not in Zeph. Zeph lists whatever that runtime is serving and lets you pick the default.
Tell the assistant which one to use
Connecting a provider makes it available; it does not make it the default. The Answer picker lives in the assistant's own settings. With nothing picked there, the assistant falls back to the first cloud provider you connected. A local runtime is not picked up automatically — set it in the Answer picker to make it the assistant's brain.
A workflow step never names a vendor. Ask agent just asks to answer, and your chosen provider serves it. Swap the provider later and every step that answers follows, untouched.
Downloading a model does not get you a brain
The on-device engines you download — Whisper, MLX Whisper, Parakeet — only transcribe. They will not answer. A model provider is either a cloud key or a local runtime. See Choose a speech provider for the dictation side.
Verified before it is stored
A pasted key is proven, not trusted: Zeph makes one probe call to the vendor and saves the key only if it answers, in the login Keychain. A local runtime is checked the same way — Test asks the daemon for its models, and a green result means it is genuinely reachable right now. Your pick is re-checked every time it is used, so a key you later remove, or a daemon you shut down, leaves the step unset rather than failing mid-run.
Next
Capabilities
How a request to answer resolves to the provider you picked — and what happens when you have not picked one.
Provider
The full catalog and the three shapes a provider takes across speech, models, and memory.
Build a workflow
Drop in Ask agent and route text through the model on its way to the cursor, the device, or the mic.