OpenWorker

by Andrew Ng

Open-source, local-first desktop AI agent that delivers finished work instead of conversation. Runs entirely on your machine with 30+ model options and 25+ native tool connectors. MIT licensed. Released July 23, 2026 (open beta).

Overview

OpenWorker is Andrew Ng’s answer to desktop AI that prioritizes privacy, local execution, and deliverables over conversation. Rather than chatting with AI, users describe tasks and OpenWorker executes them, returning finished work (documents, reports, code) that can be opened and shared immediately. Emphasizes user privacy by keeping all data local.

Release Information

  • Release Date: July 23, 2026 (open beta)
  • License: MIT open-source
  • Repository: GitHub (andrewng/OpenWorker)
  • Cost: Free (open-source)

Architecture

Local-First Design

  • Tauri 2 Desktop Shell: Cross-platform desktop framework
  • Python FastAPI Server: Local HTTP server for model inference
  • Capability Layer: Manages tool access and operations
  • Model Router: Intelligently routes tasks to appropriate models

Privacy Model: All processing, conversation history, connector tokens, and model keys stay on your machine. Secrets never enter model context.

Platforms

  • macOS: Available now
  • Windows: Coming soon
  • Linux: Potential future support

Flexible Model Support

30+ curated tool-calling models:

Frontier Models

  • OpenAI GPT-5.6 (Sol, Terra, Luna)
  • Claude Fable 5, Mythos 5, Opus 5, Sonnet 5
  • Google Gemini 3.1
  • Anthropic other Claude variants

Open-Weight Models

  • Kimi
  • GLM (Zhipu)
  • DeepSeek
  • Others

Local Models

  • Ollama (local LLM support)
  • Custom model integration via API

Native Connectors (25+)

Development & Engineering

  • GitHub
  • Terminal/Command line
  • Local file system
  • Code repositories

Productivity & Communication

  • Slack
  • Outlook
  • Gmail
  • Google Calendar
  • Linear

Project Management & Collaboration

  • Jira
  • Asana
  • Notion
  • monday.com
  • HubSpot

Local Operations

  • File read/write
  • Document processing
  • Spreadsheet operations

Typed Permission Engine

Four Risk Classifications

  1. read: Information access only
  2. write_local: Local file modification
  3. exec: Command execution on machine
  4. external: External API/service access

Five Permission Modes

  • Always allow
  • Always deny
  • Per-request approval
  • Sandboxed execution
  • Restricted operations

Shell Command Safety

  • Shell commands require explicit user approval
  • Diff previews for file modifications
  • Audit logging of operations
  • Rollback capability

Key Differentiators

vs. ChatGPT Desktop / Claude CoWork

  • Local-first architecture: Runs entirely on your machine
  • Open-source: Full transparency and customization
  • Privacy focus: No cloud dependency, no data sent to servers
  • Model flexibility: Mix and match 30+ models, including local models
  • Deliverables-first: Returns finished work, not conversation

vs. Traditional Desktop Agents

  • Privacy by default: No cloud requirement
  • Typed permissions: Granular, user-controlled access
  • Tool diversity: 25+ native connectors
  • Model agnostic: Not locked to single vendor

Use Cases

Business Intelligence & Analysis

  • Prepare customer briefs and competitive analysis
  • Generate reports from multiple data sources
  • Triage and summarize business intelligence alerts

Calendar & Schedule Management

  • Untangle calendar conflicts
  • Schedule optimization
  • Meeting preparation and note-taking

Content Generation

  • Draft emails, reports, and documentation
  • Generate marketing copy
  • Technical writing assistance

Alert Triage & Response

  • Process Slack/email alerts
  • Prioritize and summarize notifications
  • Auto-respond to routine messages

File Operations & Organization

  • Organize files across systems
  • Generate document batches
  • Archive and cleanup automation

Development Support

  • GitHub issue triage
  • Code review assistance
  • Documentation generation from code

Privacy & Security Model

  • Zero cloud dependency: Everything runs locally
  • Secret protection: API keys and credentials never enter model prompts
  • Audit logging: Track all operations performed by agents
  • User control: Granular permission management
  • Encryption: Local storage encryption available

Configuration

  • Model selection: Choose from 30+ options per task
  • Connector setup: Configure only needed integrations
  • Permission policies: Define granular access rules
  • Custom tools: Extensible architecture for custom connectors

Community & Ecosystem

  • MIT License: Allows commercial use, modification, distribution
  • GitHub Community: Open development and contribution
  • Custom Extensions: Build custom tools and connectors
  • Model plugins: Add new model providers

When to Use OpenWorker

Use for:

  • Privacy-critical tasks (no cloud upload)
  • Local-only data and operations
  • Mix of models per task
  • Open-source preference
  • Offline capability
  • Custom connector needs

Consider alternatives if:

  • Cloud sync across devices essential
  • Integrated UI/UX highly important
  • Team collaboration within app
  • Customer support from vendor needed

Competitive Position

Andrew Ng positions OpenWorker as the “privacy-first” alternative to cloud-dependent desktop AI agents. Emphasis on delivering finished work (not conversation), local execution (not cloud), and user control (granular permissions) rather than ease-of-use or cloud integration.

See Also