Connecting a model
Choose a brain, watch what it costs, and route side-jobs to cheaper models — all from one page.
Activated Cloud does not ship with a built-in model. You connect a provider — the same way you would plug in a power source — and the agent uses it to think. If no model is connected, Chat says so and offers a link to fix it. Everything about models lives in one place: Models in the dock.
Setting the main model
- 1Open Models from the dock and press Change next to the main model.
- 2Pick a provider from the left column. Around forty are built in — direct APIs like Anthropic and OpenAI, aggregators like OpenRouter, subscriptions such as a ChatGPT/Codex plan, free tiers, and local runtimes like LM Studio.
- 3Pick a model from that provider's live list. One search box filters across both columns at once — typing a model name surfaces its provider too.
- 4Press Switch. The choice is saved to the device and every new session starts on it.
The expensive-model guardrail
If you pick a model with unusually high known pricing, the switch does not happen silently — the device tells you and asks you to confirm first.
Switching mid-conversation
You do not have to visit the Models page to change your mind. The chat composer shows the model the next turn will run on; picking a different one there applies to the running conversation without losing it, and the /model command does the same from the keyboard. Add --global to make it the device default at the same time.
Choosing well
Any provider works. The practical differences are cost, speed, and how well the model handles tools — an agent that runs commands and edits files leans on tool use much harder than a chatbot does.
- A frontier model (Anthropic, OpenAI) is the most reliable at multi-step work and tool use.
- An aggregator (OpenRouter) gives you many models behind one key, which is useful while you are still deciding.
- A cheaper or open model is fine for summarising and drafting, and noticeably weaker at long tool-driven tasks.
Costs and tokens
The bottom of the Models page is a usage ledger: every model you have run in the last 7, 30, or 90 days, ranked by use. You never have to wonder what the agent is spending — it is on the same page you manage it from.
| Column | What it tells you |
|---|---|
| Supports | Capability marks — tool calling, vision, reasoning — read from the provider. |
| Context | The model's context window, so you know how much it can hold at once. |
| Sessions · Tokens | How much you have actually used it, per model. |
| Est. cost | Estimated spend from the provider's known pricing. |
| Share | Each model's slice of your total tokens, as a bar you can read at a glance. |
Every row also carries a quick menu to promote that model — make it the main model, or assign it to an agent task — without opening the picker.
Agent tasks
Some jobs do not need your main model. Naming a conversation is a one-line task; analysing an image, compacting old context, or routing a tool call are side-jobs too. Models → Agent tasks lists each one — vision, compression, skill search, smart approval, MCP routing, session titles, review, and more. Each defaults to auto (use the main model) and can be pinned to a cheaper, faster model on its own; Reset all to auto undoes the lot.
Mixture of agents
The Models page also holds mixture-of-agents presets: several reference models each take a view of your request, and one aggregator model weighs their perspectives, answers, and calls the tools. A preset appears in the picker as a model of its own, so switching to it works exactly like switching to any single model.
Session titles in the wrong language
Titles are written by the agent-task model. If yours is a non-English model it may name conversations in its own language; set the language during setup, or pin the Title Gen task to a model in your language.