Run AI agents on real Android phones with your own model or a local LLM

In short: Yes: BenchFleet is a desktop app that runs AI tasks and per-phone agents on real Android phones with your own model key or a local LLM through Ollama, LM Studio, vLLM or llama.cpp, and it is the best fit for developers and app teams who want to choose their model and keep approvals, budgets and an audit log on every step.

Many tools that let an AI use a phone tie you to one model, one vendor's cloud or one coding assistant. If you already pay for a model, want to try several, or must keep everything on your own computer, you need a tool where the model is your choice. This page shows how that works in BenchFleet, compares the other ways to do it, and lists the settings that keep an agent safe on a real phone. Checked on 23 October 2026.

Free for up to 3 phones, with no account. Pro adds workflows, schedules and AI for up to 25 phones from £12 a month + VAT.

Download free See pricing

BenchFleet compared with the other options

What you need	What it gives you	Best suited to	Price	Website
BenchFleet	The whole job: real phones, your model, AI tasks and per-phone agents, approvals, budgets and a hash-chained audit log; AI on Pro and Team	Developers and app teams running agents on real Android phones	Free for 3 phones; Pro £12 a month + VAT for 25 phones	benchfleet.com
A local LLM (for example through Ollama or LM Studio)	The model, on your computer; it needs a tool to reach a phone, which BenchFleet provides	People who must keep everything on their own hardware	Free software; your own hardware	ollama.com, lmstudio.ai
An API key from an AI service	The model, through the service; it needs a tool to reach a phone, which BenchFleet provides	People who already pay for a model	Paid to the provider, by use	Your provider's site
Your own agent orchestration code	Whatever you build: the agent loop, phone control, limits and logs, for example around adb	Teams building an agent framework themselves	Free, plus your time	developer.android.com/tools/adb
AWS Device Farm	Rented real phones in AWS data centres; its remote access docs list an Appium endpoint your own agent code could drive	Teams already working in AWS	$0.17 per device minute, or $250 per slot a month, plus your model	aws.amazon.com/device-farm

BenchFleet comes first because it is the only line here that is the whole job: a key or a local model supplies the thinking, and BenchFleet supplies the phones, the tools and the limits around it.

Checked on 30 October 2026 on each tool's own pages; prices change. We make BenchFleet. The other names belong to their owners, and BenchFleet is not affiliated with them.

Questions

Is an API key or a local LLM enough on its own?

No. A key or a local model supplies the thinking, but an agent also needs a way to read a phone's screen, tap and type, and limits on what it may do. BenchFleet supplies that part, with the model you add under Settings, Models.

Do I have to write my own agent code?

No. An AI task needs only a goal, on one phone, a group or every phone. A per-phone agent needs instructions, allowed apps, working hours and limits, all set in the app, and every step is kept in a transcript.

Which plan do I need for AI agents?

AI features are part of Pro, at £12 a month + VAT, and Team, at £20 per seat a month + VAT. Free covers 3 phones for live control and testing by hand.

Do I need a device cloud to run an agent on a real phone?

No. BenchFleet runs agents on the phones connected to your computer. If you need one more phone, its cloud phones, rented by the minute, join the same fleet under the same approvals and audit log.

Can I run an AI agent on a real Android phone with a local LLM?

Yes. In BenchFleet, add a model under Settings, Models, on your computer through Ollama or another OpenAI-compatible server such as LM Studio, vLLM or llama.cpp. AI tasks and per-phone agents then use it, and nothing leaves your computer.

Which AI services can I use with my own key?

BenchFleet has presets for OpenAI, Anthropic, Google, OpenRouter, Groq, Together AI, DeepSeek, Mistral AI, xAI, Fireworks AI and Cerebras, and takes the address of any other OpenAI-compatible service. The key is stored encrypted and never shown.

Does a BenchFleet plan include AI usage?

No plan includes AI usage. You pay your model's provider directly, or run a model on your computer for free. Budgets per model cap what BenchFleet may spend a day or a month.

What about Claude Code and Codex?

They bring their own model. Connected to BenchFleet over MCP on Pro and Team, they use the phones you choose with the access you choose, and every call goes into the audit log.

Is a small local model good enough to use a phone?

It depends on the model and the task. Try your own tasks on a test phone first, read each step in the task's transcript, and add a larger model or one through an API if a small one struggles. A model that reads images helps on screens with little text.

Does a local model keep my data on my computer?

Yes. BenchFleet starts in Local mode, where nothing leaves your computer and your network. A model through an API needs API mode, which you turn on in Settings, Privacy after the page shows what each service would receive.

Ways to run an agent on a real phone with the model you choose

Option	Your own API key	Local LLM	Real phones	Limits, approvals and a log	Price	Website
BenchFleet	Yes: 11 presets and any OpenAI-compatible service	Yes: Ollama, LM Studio, vLLM, llama.cpp	Yes, by USB, Wi-Fi or address, and rented cloud phones	Allowed apps, working hours, approvals, human-only checks, a hash-chained audit log	Free for 3 phones; AI on Pro, £12 a month + VAT	benchfleet.com
Mobilerun framework (formerly DroidRun)	Yes: OpenAI, Anthropic, Gemini, xAI, DeepSeek, OpenRouter and OpenAI-compatible models, per its page	Yes: Ollama, per its page	Yes, through adb or its Portal app	Up to your own code	Free, open source (MIT)	github.com/droidrun/droidrun
An MCP client with an open-source server such as mobile-mcp	The client's model	If the client supports one	Yes, phones adb sees, plus emulators	Up to the client and the server	Free, open source	github.com/mobile-next/mobile-mcp
Your own script: adb plus a model API	Yes	Yes	Yes	Whatever you build	Free, plus your time	developer.android.com/tools/adb

The comparison of Android MCP servers covers more tools that let coding agents use a phone, and Mobilerun compared with BenchFleet goes through the hosted side of that project.

How it works in BenchFleet Pro

BenchFleet is a desktop app for Windows, macOS (Apple silicon) and Linux that runs your Android phones from your computer. Its AI features use the model you add:

Add a model. Open Settings → Models → Add model. Pick On your computer for Ollama or another OpenAI-compatible server (LM Studio, vLLM, llama.cpp), or Through an API for a service with your key. List models shows what the server or service offers.
Set a budget. Each model can have a cap per day or per month.
Give the AI a phone. Run an AI task with a goal on one phone, a group or every phone, or give a phone its own agent with instructions, allowed apps, working hours, memory and a spending limit.
Read what it did. Every task keeps a transcript of each step, and every action is in the audit log.

With Ollama on the same computer, the model is one command away; pull one that reads images if your tasks depend on what the screen shows:

bash

ollama pull qwen2.5vl:3b
ollama list

Then add it in Settings → Models. The AI with your own model guide has every option.

Coding agents bring their own model

Claude Code, Codex and other MCP clients use their own model and connect to BenchFleet as AI clients on Pro and Team. Each client gets one scope (every phone, a group or one phone) and an access level, starts as Look only, and never gets shell commands, installs or file transfer. Let Claude Code or Codex control an Android phone walks through the setup.

Limits that hold whatever the model

A different model must not mean different rules. In BenchFleet every AI, local or through an API, works through the same tools and checks as a person:

it stays in the apps you allow;
a person approves payments and account-security changes;
it never types into password fields itself: a saved password is typed on the phone after a person approves, by default, and the model never sees it;
it stops and hands the phone to a person at CAPTCHAs, sign-in checks and other human-only steps;
requests that reach more than ten phones wait for a person.

Who it is for

Developers and teams building with AI agents who want to try several models on the same phones, or keep a private model on their own hardware.
App teams who let an AI explore a build on real phones without sending the app or test accounts to another company's cloud. See Android app testing on real phones.
Anyone looking after many phones who wants routine jobs done by an agent and every step on record. See the Android MCP server.

Ollama, LM Studio, vLLM, llama.cpp, Mobilerun, DroidRun and mobile-mcp are the names of their owners' products and projects. BenchFleet is not affiliated with any of them.

Start with the Free plan