星标 11400、fork 1500：吴恩达开源 OpenWorker

Merged summary

TL;DR - Andrew Ng open-sourced OpenWorker, a local-first, model-agnostic "AI coworker" harness that delivers finished artifacts (docs, reports, HTML briefs) rather than chat replies; it has drawn ~11,400 GitHub stars and 1,500+ forks. It matters as an open, vendor-neutral reference for the agent harness layer — tool connectivity, approvals, orchestration, permissions — which the article argues is now the real competitive asset since models are commoditizing.

• Architecture treats an Agent as an orchestrable surface = system prompt + tool set + workspace. The Python backend (coworker/) defines Code / Chat / Cowork surfaces plus a resident MyHelper; Cowork ships a deliberately minimal toolset of files / search / shell / todo.
• Approval is an engine-level mechanism, not UI polish: each tool carries risk_level and requires_approval metadata (shell writes = high + mandatory approval, read-only = low), and the TurnEngine suspends on high-risk calls pending human confirmation. Emails, calendar edits, and shell commands all gate on Approve.
• Skills reuse Anthropic's SKILL.md spec (YAML frontmatter + markdown + optional scripts) with progressive disclosure — only names/descriptions are injected at session start, with bodies pulled on demand via load_skill, so existing Claude skill packs are in principle portable.
• Model-agnostic with bring-your-own keys: OpenAI, Anthropic, Gemini alongside first-class support for Kimi, GLM, DeepSeek, Qwen, MiniMax, plus Ollama for fully local runs; 25+ connectors (GitHub, Slack, Jira, Notion, Linear, HubSpot, Gmail, Calendar) and any MCP-reachable tool, each individually permissioned. Shell execution sits behind an Executor abstraction (currently LocalExecutor with a persistent shell) with source comments reserving ContainerExecutor / VMExecutor for future sandboxing. The README points builders to aisuite as the underlying foundation.