morriszdweck/liquid-agent is a personal AI agent that runs models locally on macOS and Linux. The latest release is build-12. However, the provided FACTS do not state what changed in this version, so the changes in build-12 cannot be confirmed. The project description says it runs LFM3 1.7B and LFM3 Omni 2B locally. It supports conversation and voice/image input, command execution, reading and writing files, web search, opening apps, and browser control.
The project gathers several tasks into a single agent while running models locally. According to the README summary, it handles conversation, voice/image input, file processing, and screen control through a browser-based interface. For browser control it uses the Betterwright browser. Approval comes in two types depending on the task, and there are three safety modes. Auto 200M is a classifier that auto-approves only safe actions, and the description explicitly notes that it uses INT4. The implementation language is JavaScript and the license is MIT. This information shows the capabilities the agent offers and the approval structure, but it should be distinguished from results that demonstrate actual performance.
The scope of judgment is limited. The provided FACTS contain no performance metrics showing response speed or accuracy, and no comparison with other tools. The star count is 18, and that number alone cannot be used to judge usability or trustworthiness. The changelog for build-12 is also not given, so the extent of the release's improvements cannot be confirmed. What can currently be confirmed is the description of features and the local execution/approval structure; effectiveness and degree of improvement require separate evidence.