Motriz Docs
Website

Configure

Providers and models

Connect Claude Code, Codex, ChatGPT, and the API providers from Anthropic to Ollama Cloud, then pick models, effort, and speed.

Motriz runs your agents on the provider you choose. Any supported provider can drive any surface: quick runs, flows, chat, and the co-founder all pick their model per session or step, and a single flow can mix providers freely because the same role prompts work across all of them. Everything here is configured in Settings → Providers.

Supported providers

Providers fall into three groups: two local CLI runtimes that use subscriptions you already have, one browser sign-in, and eight API-key providers.

ProviderHow it authenticatesNotes
Claude CodeLocal claude CLI with your Claude subscription loginNo API key. The binary is auto-detected; a "Claude Code CLI path" field overrides it.
CodexLocal codex CLI with your existing codex loginNeeds both the binary and the auth file. Runs as a long-lived app-server in an isolated, app-managed runtime home.
ChatGPTBrowser sign-in with your OpenAI accountNo key to paste. The card shows your plan and email once signed in.
AnthropicAPI keyOptional base URL for Anthropic-compatible endpoints.
OpenAIAPI keyDirect API access.
Azure OpenAIEndpoint URL and API keyBoth are required. The "model" is your deployment name.
GrokxAI API keyDirect API access.
Ollama CloudOllama Cloud API keyHosted models at ollama.com — see the callout below.
OpenRouterAPI keyCatalog is filtered to models that support tool calling, because the agent loop needs tools.
Hugging FaceAccess tokenUses the Inference Providers router.
CerebrasAPI keyDirect API access.
Ollama Cloud is hosted, not local

The Ollama provider connects to Ollama's cloud service at ollama.com and requires an Ollama Cloud API key. Motriz does not run local models through it — the agent loop runs on your Mac, the inference runs on Ollama's hosted models.

Connecting a provider

Each provider in Settings → Providers is a collapsible card with a status badge: "ready" or "not set up". Unconfigured providers stay visible so the card can tell you exactly what is missing, but their models never appear in the model picker. Expanding a card shows the status detail, a setup hint, the configuration fields, and the first few available models.

CLI runtimes: Claude Code and Codex

Install and sign in once in your terminal — claude auth login for Claude Code, codex login for Codex — and Motriz picks the binaries up automatically. It launches them with your login-shell environment, so the CLIs resolve exactly as they would in a terminal. If auto-detection fails, set the path in the "Claude Code CLI path" or "Codex CLI path" field.

ChatGPT: browser sign-in

Click "Sign in with ChatGPT" and complete the OAuth flow in your default browser; the button shows "Waiting for browser…" until you finish or cancel. The flow listens on local port 1455 — if another Codex or ChatGPT sign-in holds that port, close it and retry. Sign out any time from the same card.

API-key providers

Paste the key into the provider's field and Save. Key fields are password inputs that never echo a stored value — the placeholder reads "configured — enter a new key to replace" once a key exists. To rotate a key, enter the replacement; to delete it, click Remove. Azure OpenAI needs its endpoint URL as well; Anthropic optionally takes a custom base URL.

Where credentials live

All keys and paths saved in Settings are stored encrypted (AES-256-GCM) in the local SQLite database, with the encryption key in a secret.key file beside it. There is no .env file. A subset of providers also accepts environment-variable defaults — MOTRIZ_CLAUDE_BIN, MOTRIZ_CODEX_BIN, MOTRIZ_CODEX_HOME, MOTRIZ_OLLAMA_API_KEY, MOTRIZ_OPENROUTER_API_KEY, MOTRIZ_HF_TOKEN, MOTRIZ_CEREBRAS_API_KEY, and the matching *_MODEL overrides — but values saved in Settings always win. Anthropic, OpenAI, Azure OpenAI, and Grok are configured exclusively in Settings; they have no environment-variable path at all. See Privacy and safety for the full storage model.

Provider settings lock during runs

While any run is active, every provider field is disabled and Settings shows: "Stop active agents before changing provider credentials, accounts, endpoints, or CLI paths." This keeps a single execution from switching credentials or endpoints mid-request.

Models and the picker

The model picker in the composer lists only models from connected providers, grouped per provider. If nothing is connected yet, it reads "No providers connected yet — set one up in Settings → Providers." Model ids display under their commercial names — claude-opus-5 appears as "Claude Opus 5" — and each provider's default model sorts first in its group.

Catalogs and defaults

Catalogs are fetched live wherever the provider offers a listing; a static fallback list covers the case where the live catalog is unreachable. A successful live fetch is authoritative. Codex slugs the catalog marks as hidden collect under a "More" group with a "hidden" badge.

ProviderCatalog sourceDefault model
Claude CodeLive CLI listing, always merged with a static Claude listclaude-opus-5
CodexLive from the Codex app-servergpt-5.6-sol
ChatGPTLive from the ChatGPT backend; empty until signed ingpt-5.6-sol
AnthropicLive model listingclaude-opus-5
OpenAILive listing, filtered to chat- and tool-capable familiesgpt-5.5
Azure OpenAINo live listing — the built-in list doubles as likely deployment namesgpt-5.5
GrokLive model listinggrok-4.5
Ollama CloudLive tags listingqwen3.5:397b-cloud
OpenRouterLive listing, filtered to tool-calling modelsqwen/qwen3-coder
Hugging FaceLive from the Inference Providers routerQwen/Qwen3-Coder-480B-A35B-Instruct
CerebrasLive model listinggpt-oss-120b

A brand-new work item defaults to Claude Code with the provider's default model, default effort, and Fast off.

Typed model ids

Catalogs are not an allowlist. Type any id into the search field ("Search or type a model id…") and, if it isn't in the catalog, a "Use typed model" group offers it against every connected provider. New models work the day a provider ships them, and your own Azure deployment names are typed the same way.

Vision

Image input is a per-model capability. Vision-capable models carry an image icon in the picker with the tooltip "Understands images". Every shipped Claude model accepts images; the Codex and ChatGPT GPT-5 codex family does except gpt-5.3-codex-spark; on Cerebras only gemma-4 models do; Cerebras and Grok accept PNG and JPEG only. When a model's capability is ambiguous, Motriz deliberately assumes no vision: under-claiming only blocks an image paste, while over-claiming would fail the API call mid-run. The composer enforces this at paste and send time — see Work and runs.

Reasoning effort

Effort is a per-run setting, chosen from the composer's effort pill ("Reasoning effort") and mapped to each provider's real API knob. The pill offers "Default effort" plus the levels the provider supports:

ProviderEffort levels
Claude Code, Anthropiclow · medium · high · xhigh · max
Codex, ChatGPTlow · medium · high · xhigh — plus "Ultra", which exists only for GPT-5.6 Sol
Ollama CloudTwo levels, labeled "no thinking" and "thinking"
OpenAI, Azure OpenAI, Grok, OpenRouter, Hugging Face, Cerebraslow · medium · high

Switching provider or model reconciles your selection: an unsupported effort resets to the default. If a model rejects the effort parameter at run time, Motriz retries once without it and notes "⚠ {model} rejected the reasoning effort setting; continuing without it." rather than failing the run.

Fast mode

Fast is a speed tier, deliberately separate from reasoning effort. A "Fast" pill (tooltip "Fast service tier") appears in the composer only when the selected model supports it, and fast-capable models carry a lightning icon in the picker ("Supports fast mode").

  • Codex and ChatGPT map Fast to OpenAI's priority service tier; the live catalog says which models offer it.
  • Anthropic offers fast mode as a research preview on Claude Opus 5 and Claude Opus 4.8 only — the same model at higher output speed, billed at premium provider rates.
  • All other providers, including Claude Code, have no fast tier.

Fast is kept across a model switch only if the new model supports it; a stale Fast flag on an unsupported model is silently dropped instead of failing the call.

Cost and spend

Motriz estimates USD cost per run step and rolls it up the run tree, including delegated sub-agents. Each run row shows its totals inline — 41.2k in · 3.1k out · $0.86 · 3m 12s — and the Dashboard shows "Spend · 7d", the sum over all the product's runs from the last seven days.

Claude Code reports its own real USD figure in the CLI's result output, and Motriz records it verbatim — even though execution runs on your subscription login. OpenRouter and Hugging Face are priced live from their own catalogs; the remaining providers use built-in rate tables, and a model with no known rate simply shows no cost. Treat the total as a local estimate, not a provider invoice.

There are no budgets, caps, or alerts — spend is informational only. If a provider stops responding or a model disappears from a catalog, start with Troubleshooting.

Esc

Search results

Type a feature, task, provider, or error.