Appearance
AI with your own model
AI features are part of the Pro and Team plans. The plans do not include AI usage: you choose a model and pay its provider directly, or run one on your computer for free.
Connect a model in a minute
Settings → Models → Connect AI:
- Pick one: OpenRouter (one key for models from many companies, the easiest start), Claude (Anthropic), ChatGPT models (OpenAI), Gemini (Google), or On this computer with Ollama (free, nothing leaves your computer).
- Paste your API key. Where do I get a key? opens the provider's key page. There is no address to type: BenchFleet knows where each of these services is. The key is stored encrypted on your computer and never shown again.
- Confirm one sentence: your instructions and what is on the phone screen (its text and screenshots) are sent to that provider to run the AI, and nothing else leaves your computer. This turns on API mode (Settings → Privacy, where Local mode is one click away); Details lists exactly what each service receives.
- Take the recommended model and press Test. BenchFleet asks the provider which models your key can use and picks a current one that can carry out tasks on phones and see screenshots. Test makes one tiny call to it (a fraction of a cent), so a key problem shows now and not in the middle of a task.
Then Connected ✓: open a phone and use Ask AI, or start an AI task. If the key is wrong or the provider cannot be reached, Connect AI says so in plain words; if you close it before the end, it puts Local mode back when it had turned API mode on.
One key for every model
You connect a service once. Its key works for all its models: when you start a task you pick any model the service offers, from its current list with a search box, and the model you used last is chosen for you. Settings → Models lists each connected service with Test, Change key… (a new key is used by every model of that service right away) and Remove. Models you connected before this change were joined to their service automatically, with no key to type again.
When a model call fails
A failed task says what went wrong in one sentence and what to do, with the provider's raw answer under Details: a key the service did not accept, no credit left, a model the service no longer offers, too many requests at once, or the service could not be reached. With OpenRouter, a model company may refuse the key you saved for it in your OpenRouter account (OpenRouter's Integrations, "bring your own key"): OpenRouter then passes your request on with that key, and the message tells you to remove or replace it at openrouter.ai/settings/integrations, or to pick a model from another company.
More choices (Advanced)
Under Advanced on the Models page:
- Add a model by hand for any other AI service with an API (presets for Groq, Together AI, DeepSeek, Mistral AI, xAI, Fireworks AI and Cerebras, or the address of any other OpenAI-compatible service) or an AI server on your own network (LM Studio, vLLM, llama.cpp, Ollama on another computer). List models shows what the service offers.
- Spending limits per model or for all of them, per day or month.
- Defaults: which model the command palette, AI steps in workflows and screen recognition use.
What AI can do
- AI tasks: write a prompt for one phone, a group or every phone; read every step in the task's transcript. AI works only on phones whose AI and automation switch is On (On, Off or Paused).
Ask AI on a phone's page
Open a phone from the Fleet page. Under the AI and automation switch in the right-hand column, Ask AI takes a prompt for that phone: type what you want, pick the model (any model of a connected service; the one you used last is chosen), then Start. It may open any app; list the ones it should leave alone under Apps it may not open. If a task stops at an app on that list, Allow (the app) and try again runs it once more with that app allowed.
Opening a phone puts you in control of it, and AI waits while you are, so the button reads Release control and start: the phone is given back and the task begins. You watch its steps appear in the panel, with Pause and Stop, and Open in Runs shows the same run with everything else. The last few tasks on the phone are listed under it; a finished task offers Save as workflow….
When a task cannot run, the panel says why and offers the fix: add a model (Settings → Models), switch to API mode for an internet model (Settings → Privacy), turn AI and automation On when it is Off or Paused, or release control. A model you added but cannot pick is listed with the reason, for example API mode is off or this model cannot call tools. More options… opens the full AI task dialog for simulation, limits and guidance. Runs → AI task… does the same for a group or every phone.
- Agents: give a phone its own agent with instructions, allowed apps, working hours, memory and limits (AI agents per phone).
- AI clients: let Claude Code, Codex or any MCP client use your phones through BenchFleet (Settings → AI clients; Connect an AI client).
- Plain-language commands: type a sentence in the command palette (Ctrl+K).
- Workflow repair: after an app update breaks a workflow, the AI proposes a fix you can test before it is used.
- Why a run failed: on a failed workflow run, Explain in plain words asks the model you choose for a short summary of what went wrong (Workflows and schedules).
- Save as workflow: turn a finished AI task into a workflow (Explore with AI, replay without a model).
What AI cannot do
AI uses the same tools and checks as everyone else: only the apps you allow, a person approves payments and account changes, it never types into password fields, and it stops and asks for a person at CAPTCHAs, sign-in checks and other human-only steps.