OpenWorker: Andrew Ng’s Open‑Source AI Coworker Hits 11K Stars

In May Tencent Cloud announced WorkBuddy, an AI office companion, and in July Andrew Ng released OpenWorker, an open‑source, locally‑first, model‑agnostic AI coworker that shares the slogan “doesn’t just chat with you, but delivers finished work”.

What OpenWorker Is

One‑sentence definition: an open‑source AI colleague that lives on your computer.

Example: ask it to prepare a Northwind renewal call briefing; it fetches HubSpot records, reads email threads, and returns an HTML slide deck with data charts—a finished product, not a to‑do list.

Four Key Design Points

Deliverables, not chat. Generates documents, spreadsheets, reports, web pages as files; can be @‑mentioned in Slack and reply as a post.

Model‑agnostic. Supports OpenAI, Anthropic, Gemini and explicitly lists Kimi, GLM, DeepSeek, Qwen, MiniMax; can run fully offline with Ollama.

25+ connectors + MCP. Integrates GitHub, Slack, Jira, Notion, Linear, HubSpot, Gmail, Google Calendar, etc.; each connector can be toggled individually.

Approval‑gate for risky actions. Sending email, modifying calendar, executing shell commands all require explicit user approval.

Architecture Overview

The source code resides under coworker/ (Python backend). Key observations:

Agent is the surface, not the model. agents/ defines Code, Chat, Cowork agents and a persistent MyHelper; Cowork’s toolset is limited to files, search, shell, todo.

System prompts embed engineering discipline. Cowork instructions force a todo_write plan before any action and prohibit multi‑line heredocs; scripts must be written to file first for auditability. Artifacts are returned as links.

Approval is engine‑level. Each tool carries risk_level and requires_approval metadata; the TurnEngine pauses high‑risk tools and routes them through an approval path.

Skills reuse Anthropic standard. skills/base.py loads SKILL.md files (YAML front‑matter + markdown + optional script) with progressive disclosure; only skill names and descriptions are injected at session start.

Executor leaves sandbox backdoor. Shell commands run via an Executor abstraction; currently a LocalExecutor maintains state across commands, with comments indicating a future ContainerExecutor / VMExecutor for isolation.

Comparison with Tencent WorkBuddy

Both target AI‑augmented office work, but their roadmaps diverge: Tencent bets on a tightly integrated ecosystem (“platform family bucket”), while OpenWorker offers an open, self‑hosted harness that gives users control over data and models.

Takeaways

The README’s headline—“If you want to build your own agent harness rather than use ours, start with aisuite”—highlights that the core asset of an agent is the harness (tool integration, approval flow, task orchestration, permission control), not the underlying model.

With Chinese models now listed alongside OpenAI and Claude, the model layer is becoming commoditized; the competitive edge will be the quality of the harness.

https://github.com/andrewyng/openworker
https://medium.com/@sudarshan-koirala/openworker-andrew-ngs-open-source-ai-coworker-that-delivers-finished-work-84509e57330c