MiniMax rolls out Pocket and “Computer Use” to push its desktop agent beyond the command line
MiniMax has expanded its desktop AI agent from a command-line helper into a screen-operating worker, betting that real productivity gains require control over local apps, internal systems and graphical interfaces that APIs can’t reach.
The company on Tuesday launched two desktop updates: Pocket, a beta feature that lets users trigger tasks from mainstream messaging apps such as Feishu, WeChat, WeCom and Slack, and “Computer Use”, which enables the agent to watch the screen and operate a user’s mouse and keyboard like a human.
The release extends MiniMax’s agent stack one week after it introduced MMX-CLI, positioning chat as a universal entry point while shifting more execution onto the user’s own computer. For investors watching China’s agent race, the message is clear: differentiation is moving from model demos to end-to-end task completion, reliability and permission controls.
Expanding access shifts agents from API tasks to desktop workflows
Pocket turns IM threads into a remote control for a user’s computer: send an instruction in chat, the agent runs the workflow locally and returns results to the same conversation. MiniMax showcased scenarios including finding a local PDF on a desktop and sending it back to the chat, as well as reading a folder of resumes, selecting top candidates, drafting interview questions and writing the output into a Feishu document before returning a link.
Computer Use targets the long tail of work that sits behind GUIs, including local design tools, internal reporting systems and operating-system settings. In one example, the agent navigated system settings to disable a screen saver option, then opened a Pocket client to configure a daily scheduled task and returned a screenshot—steps that MiniMax said were not accessible through public command-line interfaces.
Re-architecting toolchains aims to improve precision and execution quality
Rather than exposing a single “computer tool” that clicks by pixel coordinates, MiniMax said it split desktop operation into four tool domains: Desktop Control, Window Manager, Browser Engine and Clipboard. The goal is to route each step to the most reliable control surface—system APIs for window management, DOM selectors for browser elements, and clipboard operations for cross-app data transfer—reducing dependence on fragile visual clicking.
The company said the matrix is layered on top of multiple platform CLIs, including lark-cli, wecom-cli and mmx, plus Bash and file-system tools, totaling more than 60 tools. The agent selects the execution path based on task type, an approach intended to improve determinism in real workflows.
Normalizing vision inputs is designed to make screen control portable across devices
MiniMax said desktop vision breaks when screen sizes and resolutions vary, from 720p displays to external 4K monitors, because the same screenshot can overwhelm a model with detail or blur critical UI elements. Its approach standardizes interaction in two ways: the model outputs a 0-to-1 relative coordinate rather than absolute pixels, and the system converts that to device-specific coordinates at runtime.
It also applies adaptive screenshot processing, capturing at physical pixel fidelity and then dynamically resizing to the model’s visual input limits to preserve legibility without exceeding capacity. The company framed this as necessary for consistent UI targeting across hardware.
Adding verification and IM-based permissions targets risk in unattended automation
To reduce failure rates in long, multi-step workflows, MiniMax said each desktop action is followed by an automated screenshot verification step. If the expected state isn’t reached, the agent enters a diagnostic flow and attempts alternatives—such as switching from mouse interaction to keyboard shortcuts—before reporting where it is stuck rather than continuing blindly.
Because IM-triggered automation can run while users are away from their computer, MiniMax embedded permission gates inside chat. When the agent is about to perform sensitive actions such as file deletion or system changes, it pauses and sends an authorization request to the IM conversation. Feishu and Slack can display interactive cards, while platforms without interactive components rely on text-based confirmation. Users can also interrupt tasks mid-execution.
MiniMax said Pocket and Computer Use remain early-stage capabilities, with more work needed on recognition accuracy in complex interfaces, stability in long tasks and generalization to new software.
Related Coverage:
MiniMax Debuts ‘Self-Evolving’ M2.7 Model to Rival Global Heavyweights
MiniMax Blows Past Estimates as Global AI Push Drives 159% Revenue Surge