Repository navigation
AutoCua 3.0.0: most of AutoCua now runs in Rust, local models with Ollama, Gemini 4 Carbon and Hydrogen
LatestMost of AutoCua now runs in Rust
Since 0.2.0 most of the agent code has moved from Python to Rust. Rust is now the largest language in the repository (about 45%, up from 23% in 0.2.0). It still ships prebuilt in the wheel, so you never need Rust. Each part was compared with the Python it replaced, case by case, before the switch.
- Computer use: one Rust agent now runs on macOS, Windows and Linux in place of three Python agents. Each system keeps its own prompts. Windows picks up what its copy had fallen behind on, such as the "Pure vision first" hint and the newer URL finder for site skills.
- iPhone agent: the agent, its controller (taps, typing, scrolls, opening apps, the shell, the video player, notes and the todo list) and its tool registry.
- LLM layer: the manager every agent sends its requests through, the Responses route and all eleven model providers.
- Shared tools: web search, the shell, the browser sub-agent and the Android sub-agent are now one copy each for every system, in
AutoCua/universal_tool. They replace six copies of the web tool and fifteen copies of the others. - Also in Rust: the tool registries, notes and todo lists for macOS, Windows and Linux, the vault, memory compression and the app's chat memory.
- The shell tools of computer use and the iPhone agent now run in the coder's sandbox. The old Python sandboxes are gone.
- The screen and phone readers, the desktop controllers and the app itself are still Python.
The version jumps from 0.2.0 to 3.0.0 to mark this move. There is no 1.x or 2.x.
New provider: Ollama, models on your own machine
provider="ollama" runs the agents on a model on your machine, with no key.
- It reads
OLLAMA_HOSTthe way Ollama's own client does.OLLAMA_API_KEYis only for an Ollama that sits behind a key. - A model is checked once for tool calling and vision, and the run stops with a reason when it lacks what the agent needs. Its context window is checked too.
- Three models are in the table, tested end to end:
ornith-1.5:35b,qwen3.6:35b-a3b-nvfp4andqwen3.8:27b. Any other pulled model with tool calling works too. - The app lists Ollama with no key saved, shows its models in the chat picker, and has a Settings row that says whether Ollama answers. Telegram offers it while it runs.
- Web search goes through the headless browser agent, as on Together, Cerebras, Requesty and AWS.
Models
- Gemini 4 Carbon and Gemini 4 Hydrogen on Google (AI Studio and Vertex), OpenRouter, Perplexity and Requesty.
- GPT-6.1 Sol and GPT-6.1 Terra on the direct OpenAI provider.
- 75 models across 11 providers in total.
The app
- The mode "Web use" is now called Browser.
- One
.envfor every key. Settings and the Telegram bot now write the same.envyou edit by hand. An oldapi_key.txtmoves into it once, by itself. - iPhone pairing: the bundle ID is optional and defaults to your own team's (
com.AutoCua.<team ID>), so pairing works on any Apple account with nothing to set. Settings has a Bundle ID box if you want your own. - iPhone pairing runs one at a time and has a Stop button.
- First launch: when GitHub's release link fails, the app's window download now comes through GitHub's API instead.
Browser agent
- macOS 27: the agent can use your own Chrome again. The system keeps a Chrome started by another program out of its profile, so Chrome is now started through the system launcher.
- One tab, not two, when the agent starts in a Chrome that has no window open.
- A running Chrome is still found when its profile lock cannot be read.
Web search
- One web search tool for every agent.
- Each provider's library loads at its first search, so a missing one no longer stops an agent from starting.
- Ctrl+C during the browser fallback stops the browser agent too.
- Windows: a failed search is retried 3 times and reported as an error. It used to come back as a success.
- Windows: a status line with an emoji can no longer fail the search when the output cannot show it.
Other fixes
- Android: a login saved under the app's name (for example "Instagram") is now found when that app is in front.
- iPhone: in parallel simulator runs the agent now finds its own notes.
- Shell runs: leftover output after a command finished no longer turns it into an error. Thanks to @kayaal34 (#109).
Changes you may need to act on
- Mode names: in
main.pyand the terminal, usemode="browser"andmode="coder". The old names "web use" and "shell use" are gone. - Browser agent path:
python -m AutoCua.browser.agentreplacespython -m AutoCua.web.agent. - Vault: the section for site logins in
credentials.jsonis now"browser". It was"web". - Old chats: chats saved under the old Web use and Shell use modes do not reopen under the new names.
- iPhone: a phone paired under the old bundle ID
com.AutoCuaasks for pairing once more.
Install or upgrade
pip install -U autocua
Wheels for macOS (Apple Silicon), Windows x64 and Linux x86_64.
New Contributors
Full Changelog: v0.2.0...v3.0.0