Andrew Ng has shipped OpenCoworker, a desktop AI agent that lives on your machine and actually does work alongside you. Built on top of aisuite—Ng's lightweight Python library for LLM orchestration—the app launched with downloads available for macOS (Apple Silicon, macOS 13+) and Windows (10/11 x64). Unlike cloud-based assistants that farm your data to remote servers, OpenCoworker keeps everything local by default.

What It Actually Does

OpenCoworker goes beyond chat. The agent can read files on your system with permission to gather context, send messages through Slack and email, and produce real deliverables like PDF reports, Word documents, and spreadsheets. It also handles scheduled automations—think a daily news briefing delivered to your inbox every morning without you lifting a finger.

aisuite: The Foundation

The underlying aisuite library provides two layers of abstraction. First, there's the Chat Completions API—a unified, OpenAI-style interface that works across OpenAI, Anthropic, Google, Mistral, Hugging Face, AWS, Cohere, Ollama, and OpenRouter. Developers swap providers by changing a single string in their code. Second, the Agents API adds tool calling, multi-turn loops, and prebuilt toolkits for file operations, git, and shell execution.

Bring Your Own Keys—or Go Fully Local

The app requires an API key from your preferred provider (OpenAI, Anthropic, or Google), but Ng's team also built in first-class Ollama support. That means you can run the entire stack locally on your own hardware—no subscription fees, no data leaving your machine. This local-first approach is a direct answer to privacy concerns that have plagued cloud AI assistants.

The Agents API Under the Hood

For developers building custom agent harnesses, aisuite exposes tool calling as a one-liner: pass plain Python functions and it generates schemas, executes calls, and feeds results back to the model. Tool policies let you enforce approval workflows or block specific operations entirely. State stores support persistence across sessions via in-memory, file, or Postgres backends.

MCP Support Out of the Box

OpenCoworker natively supports Model Context Protocol servers, so any MCP tool can be handed to a model without custom boilerplate. A filesystem server example shows how simple it is: pass the MCP configuration and the agent gains access to read/write operations on designated directories.

Key Takeaways

  • OpenCoworker runs fully locally via Ollama—no API keys needed if you want total privacy
  • Built on aisuite with unified support for 10+ LLM providers including Anthropic, Google, and Mistral
  • Ships as a downloadable app for macOS and Windows today
  • Source code under platform/ serves as a reference implementation for building your own agent harness

The Bottom Line

Ng's OpenCoworker isn't just another chatbot—it's a serious attempt to put AI to work on real desktop tasks while keeping data in users' hands. The local-first design and MIT licensing make it worth watching, especially if you've been burned by cloud AI privacy nightmares before.