OpenDazi
A desktop AI agent that works with the files on your machine. I wrote nearly all of it; it is built with Electron and TypeScript, and private for now.
- Chat about a project, with the plan and tool calls visible, and keep steering while it works
- Board: every card is a task with its own conversation, materials and results. It runs in the background and comes back for review
- Workflows: once a task passes its checks it can be saved as a workflow, then rerun on new inputs
- Meetings: local or cloud transcription into a transcript, a summary and a mind map, with action items sent to the board
- Memory: preferences and project conventions picked up from conversations, which you can review and edit
- Scheduled tasks that run a prompt on a timetable, system audio capture on macOS, and Windows and Linux builds for internal testers
verl-agentic
An RL framework for Android GUI agents. Private for now.
Rather than maintain a fork, it is being rebuilt as a plugin on top of upstream veRL. The plugin focuses on step-level learning signals, which is where GUI tasks are hardest:
- grouping steps by screen state, so an action is compared against others taken from the same state
- forking rollouts from environment snapshots instead of replaying from the start
- self-distillation from a teacher that sees privileged information
langextract-vllm
A provider plugin that lets Google’s LangExtract run structured extraction against a vLLM server.
- Source: wuli666/langextract-vllm, 30 stars
- Listed in LangExtract’s community provider registry (google/langextract#244)