AI Providers

Bring your own vision models for agent runs and description grounding.

Yoqa is BYO — you paste API keys or use local CLIs. No Yoqa cloud credits are required for Phase 1 local runs.

Where to configure

Desktop app → Settings → Providers:

1. Add a provider.
2. Paste the key (or complete CLI login where applicable).
3. Validate — Yoqa probes models / auth.
4. Set a default provider (and model when the list is available).
5. Confirm with yoqa status (provider: …).

Supported providers

Provider	Auth	Notes
Anthropic	API key	Claude vision via AI SDK
OpenAI	API key	GPT vision via AI SDK
Groq	API key	Fast OpenAI-compatible vision models
Google	API key	Google AI Studio / Gemini
Google Vertex	GCP credentials	Vertex AI Gemini
Antigravity (agy)	CLI login (+ optional AI Studio key)	Vision via agy --print
OpenCode	Zen API key	Vision runs against hosted Zen (opencode.ai/zen/v1). CLI login alone is not enough — local opencode serve has no OpenAI-compatible /v1. Defaults to the free mimo-v2.5-free model
Cursor	Cursor Agent CLI (+ optional key)	Vision via cursor-agent --print
Grok (xAI)	API key	@ai-sdk/xai
Custom	Optional key + Base URL	Any OpenAI-compatible gateway (e.g. local Ollama /v1)
Codex / Claude CLI	CLI where supported	Additional local agent backends

What needs a provider?

Capability	Needs provider?
yoqa screen / coordinate actions	No
Description grounding (-d)	Yes
Autonomous case runs (Yoqa agent)	Yes
Apps / cases / flows CRUD	No

Tips

• Prefer models that accept image inputs — text-only models fail vision decide.
• For OpenCode, paste a Zen API key from opencode.ai in Settings → Providers. Yoqa calls hosted Zen at opencode.ai/zen/v1. A local opencode serve does not expose OpenAI /v1/chat/completions, so CLI login alone will not authenticate vision. New providers default to mimo-v2.5-free; most other free Zen models are text-only and reject screenshots — switch to mimo-v2.5-free (or a paid multimodal model on Zen with billing) if a run reports “No endpoints found that support image input”.
• Custom base URLs are ideal for private gateways; leave the API key empty only if the host allows it.

How the agent works

Where vision is used in the loop

Quickstart

First run with a provider configured