OpenWorker

OpenWorker is a 100% open-source AI coworker that runs on your desktop and delivers finished work — polished documents, Slack replies, updated calendars, triaged inboxes — rather than just chat responses. Created by Andrew Ng and released under the MIT license, it is currently in open beta on macOS (Apple Silicon), with a Windows build available but not yet code-signed. The project is built on aisuite, Ng's lightweight Python library for unified LLM access.

What It Is

OpenWorker is a local-first autonomous agent harness that sits on your machine and orchestrates multi-step tasks across 25+ connected tools. You describe the outcome you want — "prep me for the Northwind renewal call," "untangle my Tuesday," "triage the checkout alert" — and OpenWorker breaks the task into steps, works across your desktop and connected apps, and returns a finished deliverable. Before any consequential action (sending an email, posting to Slack, running a shell command), it pauses and asks for your approval. The agent loop, conversations, connector tokens, and model keys all live locally; the only cloud component is a small OAuth broker for one-click tool connections.

Architecture and Local-First Design

Under the hood, OpenWorker is a three-layer stack:

• Native desktop shell — a Tauri-based GUI (React + Rust) that supervises the server process and renders the chat and artifact views.
• Local agent server — a Python backend (coworker/) that runs the agent engine, model providers, connectors, MCP client, memory, and automations.
• Your keys, your machine — all data leaves the machine only through the model API and integrations you explicitly configure.

The app auto-updates, so fixes reach installs quickly. It can also be run from source with Python 3.10+, Node 20+, and the Rust toolchain.

Model Flexibility

OpenWorker does not bundle or lock you into any model. Supported providers out of the box include OpenAI, Anthropic, Google Gemini, DeepSeek, Kimi (Moonshot), Qwen, MiniMax, Mistral, Grok (xAI), GLM (Z.ai), Together AI, and Fireworks. For fully local inference, Ollama is supported. You can switch models per task or even mid-conversation. A curated model list marks providers verified for reliable tool-calling.

Tool Integrations and Trigger Modes

OpenWorker ships with 25+ connectors covering the most common workplace tools:

• Communication — Slack, Gmail, Outlook
• Calendar — Google Calendar
• Project management — Jira, Linear, Asana, Notion
• CRM — HubSpot, Attio
• Files and storage — Google Drive, Dropbox, Box
• Dev tools — GitHub, terminal/local files
• Protocol — any MCP-compatible tool plugs in with per-tool control

Beyond interactive chat, OpenWorker supports trigger-based automations: a Slack mention, an incoming email, or a scheduled time can kick off a task automatically. Unattended runs park approval requests in an inbox rather than acting autonomously.

Pre-Built Use Cases

The app ships with four ready-made "coworker" configurations:

• Sales — researches accounts, preps meetings, drafts follow-ups; integrates with HubSpot, Email, and Calendar.
• Executive Assistant — manages calendar conflicts, triages inbox, drafts replies; integrates with Email, Calendar, and Slack.
• Marketing — tracks spend and attribution, generates reports; integrates with HubSpot and Slack.
• Ops On-call — triages alerts, digs through runbooks, drafts incident notes; integrates with Slack and GitHub.

Update: v0.1.6 Open Beta

The latest release is v0.1.6, published on July 23, 2026. The repository was created on July 20, 2026 and has accumulated over 3,300 GitHub stars and 450 forks within days of launch, reflecting rapid early community interest. The project self-describes as "open beta: fully usable, updates itself, and we're actively polishing rough edges." Windows builds are available but code-signing is still in progress, so SmartScreen warnings appear on Windows installs. The project originated inside the aisuite repository before moving to its own home.