AutoCua 0.2.0: the coder rewritten in Rust, Coder mode in the app, safer Telegram pairing

[FunctionFreak]

The coder is now written in Rust

The coder, the agent behind Coder mode and computer use's coding sub-agent, has been rewritten in Rust. One copy now runs on macOS, Windows and Linux in place of three Python copies. It ships prebuilt in the wheel next to the web and Android agents, so you still never need Rust.

• It writes down what it concluded (insight) and the goal of each step (next_goal), not its reasoning.
• Bigger tasks start in planning mode: read-only exploration, a written plan with exact path:line anchors, and a ToDo list.
• Subagents replace minions and come in two modes. workflow maps the code for the coder, and critic reviews an edit and tries to prove it wrong. Each runs as its own process.
• When a request does not say where something lives, the coder sends a workflow subagent to find it instead of searching through files itself.
• Its file tool is now called read (it was view).
• Run it on its own with python -m AutoCua.coder --task "..." --provider ... --model ..., or from main.py with mode="coder".

Coder mode in the app

• Shell use is now called Coder.
• A Coder chat remembers the whole conversation. Every request is numbered, and a short note between requests records whether the last one finished, was stopped or failed, so the next request picks up from there.
• Subagents appear live as subagent bots, in Coder mode and in computer use's coder box.
• Windows: Coder mode no longer hangs at start in the desktop window.

Coder mode thinks deeper

• In Coder mode the model thinks one level deeper than in the other agents (for example low becomes medium). It can write up to 32K tokens per reply and waits up to 10 minutes per call. The other agents are unchanged.
• Claude Opus 5 and Sonnet 5 (Anthropic, OpenRouter, Requesty) and Claude on Amazon Bedrock now really think in Coder mode.
• Each model's Coder-mode settings sit next to its normal ones in the provider tables, and model_list.txt has a new coder column.

New provider: Requesty

Requesty works in the app, the browser agent and the Telegram bot, with native tool calling and its own key row in Settings. Thanks to @Thibaultjaigu (#84).

Models

• OpenRouter adds gemini-4-argon, gpt-6.1-terra, deepseek-v4.1-flash-1017, minimax-m3.1-flash and muse-wonder. It drops gemini-3.1-pro, muse-spark-1.3, kimi-k3 and kimi-k3.1-pro.
• Requesty adds claude-sonnet-5.5, claude-opus-5, claude-opus-5.5, gemini-4-argon, grok-4.7, particle-1, minimax-m3.1-flash and deepseek-v4.1-flash-1017, with prompt caching.
• Google and Perplexity: gemini-4-argon replaces Gemini 3.1 Pro. Perplexity drops grok-4.3.
• 60 models across 10 providers in total. Every new model was tested live. See model_list.txt.

Security

• Telegram pairing (Security: Telegram bot hands device control to any chat when the owner line is missing #85): the bot no longer gives control to the first chat that messages it. Connect shows a one-time code, in the banner and in Settings, Remote Connection. Only a private message with that code pairs the bot, and until then it obeys nobody. Disconnect clears the token, the owner and the code.
• A crash or two Settings saves at once can no longer empty api_key.txt, and the file is now readable by you only.
• Fixes from code scanning: the desktop window's session token is no longer written to the backend's output, the login redirect stays on the same host, text typed in the terminal box is never treated as markup, error replies carry only the error's own message, and the Android Wi-Fi check sends from one network interface only.

Setup

• Windows setup no longer installs the Interception driver, asks for Administrator rights or reboots. It reminds you that answering UAC prompts needs the driver (README, "Answering UAC prompts on Windows").
• The setup scripts never download Python. Without a suitable Python they point you to python.org on macOS and Windows, or to your distribution's package on Linux.
• iOS setup lives in the app only (Settings, Connect Device, iPhone). The standalone browser page is gone.

Other fixes

• Browser agent: page-bridge fixes for scripts that contain U+2028/U+2029 and for wait timeouts.
• Gemini's cached tokens are now counted.
• Dependencies: Pillow 13.0.0 on Windows, Electron 44.9.1.
• Links now point to github.com/ComputerUseAgent/AutoCua.

Install or upgrade

pip install -U autocua

Wheels for macOS (Apple Silicon), Windows x64 and Linux x86_64.

New Contributors

• @Thibaultjaigu made their first contribution in Requesty provider #84

Full Changelog: v0.1.9...v0.2.0

---

This discussion was created from the release AutoCua 0.2.0: the coder rewritten in Rust, Coder mode in the app, safer Telegram pairing.