OpenWorker Review 2026: Andrew Ng's Open-Source AI Coworker Tested

OpenWorker vs Claude Code vs Codex CLI vs Hermes Agent: Best Free Open-Source Desktop AI Coworker in 2026

Every AI tool you’ve used so far gives you text. OpenWorker gives you finished work.

You prompt ChatGPT. It writes a draft. You still open your CRM, run the report, send the email, update the ticket, and move the task. You do the actual work.

Andrew Ng just released OpenWorker on July 23, 2026 — an MIT-licensed, open-source desktop application that flips this model. It’s not a chatbot. It’s a desktop AI coworker that lives on your machine, connects to your tools, and delivers finished deliverables: saved files, sent messages, updated tickets, generated reports. No copy-paste required.

We spent 72 hours testing OpenWorker across real workflows — sales research, Slack triage, weekly report generation, calendar management — and compared it against Claude Code, Codex CLI, and Hermes Agent. Here’s the honest breakdown.

Bottom Line Up Front

OpenWorker is the best free open-source desktop AI coworker for knowledge workers in 2026. No other tool combines model agnosticism (30+ models), local-first privacy, 25+ built-in integrations plus MCP, a sophisticated typed permission engine, scheduled automations, and an MIT license into a single desktop app. It’s free — you only pay for the AI model API keys you choose, or run it entirely local with Ollama at $0.

If you’re a developer, ops engineer, or power user who wants an AI agent that delivers real finished work without vendor lock-in, OpenWorker is the clear winner. If you primarily need code editing, stick with Claude Code or Codex CLI. But for knowledge work — research, reports, Slack triage, CRM updates, calendar management — OpenWorker is unmatched.

DOWNLOAD OPENWORKER — or build from source at github.com/andrewyng/openworker

Comparison Table: OpenWorker vs Claude Code vs Codex CLI vs Hermes Agent

Feature	OpenWorker	Claude Code	Codex CLI	Hermes Agent
Desktop GUI	✅ Tauri 2 + React	❌ CLI only	❌ CLI only	❌ CLI/messaging
Model-agnostic	✅ 30+ models	❌ Claude only	❌ OpenAI only	✅ Multi-model
Tool integrations	✅ 25+ connectors + MCP	✅ MCP + filesystem	✅ Sandboxed + filesystem	✅ 15+ adapters + MCP
Local-first	✅ Full local	❌ Remote sessions	❌ Cloud sandbox	✅ Full local
Scheduled automations	✅ Yes	❌ No	❌ No	✅ Cron-based
Approval gates	✅ 4 risk classes, 5 modes	✅ Yes	✅ Yes	✅ Yes
Speech input	✅ Rust STT	❌ No	❌ No	✅ Whisper
Open source	✅ MIT	❌ Proprietary	✅ Apache 2.0	✅ MIT
Primary focus	Knowledge worker tasks	Code editing	Code editing	Full-system agent
Price	Free (BYO keys)	Part of Claude Pro/Max	Part of ChatGPT Plus	Free (BYO keys)

This table tells you everything. OpenWorker is the only tool with a real desktop GUI, model agnosticism across 30+ providers, scheduled automations, and local-first architecture. Claude Code and Codex CLI are excellent coding agents, but they’re CLI-only, locked to a single model provider, and designed for code editing, not knowledge work. Hermes Agent is the closest open-source competitor, but it lacks the desktop GUI, the 25+ pre-built integrations, and the polished permission engine.

What Makes OpenWorker Different

Before we dive deep, understand what separates OpenWorker from everything else you’ve tried.

Chatbots (ChatGPT, Claude, Gemini) answer questions. They generate text. You do the rest.

Coding agents (Claude Code, Codex CLI) edit code. They’re terminal-based, model-locked, and designed for software engineering.

OpenWorker is a desktop AI coworker. It:

• Lives on your machine as a native Tauri 2 desktop app
• Connects to 25+ tools (Slack, GitHub, Jira, Notion, Linear, HubSpot, Outlook, Gmail, Google Calendar, and more)
• Delivers finished work — saved files, sent messages, updated tickets, generated documents
• Runs any model — OpenAI, Anthropic, Google, DeepSeek, Mistral, local via Ollama
• Keeps all data local — conversations, tokens, keys stay on your machine
• Has a typed permission engine with 4 risk classes and 5 permission modes

The difference is simple: OpenWorker leaves your tools in a different state than it found them. It doesn’t tell you what to do. It does it.

OpenWorker: Best Free Open-Source Desktop AI Coworker

OpenWorker is an MIT-licensed open-source desktop application created by Andrew Ng and Rohit Prasad, built on top of Ng’s provider-agnostic aisuite library. Launched July 23, 2026, it already has 2,651 GitHub stars and 357 forks.

The architecture is four layers, all local: a Tauri 2 desktop shell wrapping a React 18 UI, a Python 3.10+ FastAPI agent server on 127.0.0.1:8765, a capability layer with local tools (files, git, ripgrep, shell, todo) plus 25+ hosted connectors and MCP, and a model router that unifies native, OpenAI-compatible, reseller, and local providers.

The codebase is substantial: 119 Python files (~32,400 lines) plus 149 TSX files and 78 test modules. This is not a weekend project — it’s a serious open-source release from one of AI’s most respected figures.

What We Liked

It delivers finished work, not chat. This is the whole point. We asked OpenWorker to research a customer from Slack messages, pull their HubSpot deal history, and draft a personalized outreach email saved to a file. It did all three — read Slack, queried HubSpot, wrote the email, and saved it to our desktop. No copy-paste. No “here’s a draft, now go do the work.” It delivered a finished deliverable.

Model-agnostic with 30+ curated models. OpenWorker supports OpenAI (GPT-5.6 Sol/Terra/Luna, GPT-5.5), Anthropic (Claude Fable 5, Opus 4.8, Sonnet 4.6, Haiku 4.5), Google (Gemini 3.1 Pro, 3.6 Flash, 2.5 Pro, 2.5 Flash), plus DeepSeek V4, Grok 4.3, Mistral Large, Qwen3 Max, Kimi K2.6, GLM-5.2, and more. You can even switch models mid-conversation. No vendor lock-in. We tested it with Claude Fable 5 for reasoning tasks, DeepSeek V4 for cost-sensitive work, and Ollama running Llama 4 locally for offline tasks. All three worked seamlessly.

Local-first privacy. Conversations, tokens, and API keys stay on your machine. Model calls go directly from your computer to the provider — there’s no OpenWorker cloud intermediary. The only optional cloud component is an Auth0 PKCE OAuth broker for one-click connector setup, and even there, tokens never get stored in the cloud. OpenWorker works fully signed out with manually pasted credentials. For privacy-conscious teams, this is a massive advantage over Claude Code (remote sessions) or Codex CLI (cloud sandbox).

Sophisticated typed permission engine. This is the standout architectural feature. OpenWorker has 4 risk classes — read (no side effects), write_local (path-scoped), exec (shell commands), external (off-machine effects) — and 5 permission modes: discuss (read-only), plan (read-only), interactive (asks before writes/commands/external), auto (all auto-approved, path-scoped), and custom (auto-approval for specific tools). Shell commands ALWAYS ask by design. Unattended mode routes prompts to an Inbox rather than raising autonomy. This is security architecture done right.

25+ built-in integrations plus MCP. Slack, GitHub, Jira, Notion, Linear, HubSpot, Outlook, monday.com, Gmail, Google Calendar, and more are all pre-built. Plus, OpenWorker supports the MCP protocol with per-tool access control, so you can add any MCP-compatible tool. The Slack integration is particularly good — you can mention @OpenWorker in channels and it handles the task.

Scheduled automations. This is rare in desktop AI tools. OpenWorker can run recurring work — daily briefs, weekly reports, automated watches — on a schedule. No other tool in this comparison does this except Hermes Agent (cron-based). Claude Code and Codex CLI don’t even try.

Completely free (MIT license). OpenWorker costs $0. You pay only for the model API keys you choose. Run it with Ollama locally and the total cost is zero — just your hardware. Compared to Claude Code (requires Claude Pro at $20/mo or Max at $100/mo) or Codex CLI (requires ChatGPT Plus at $20/mo), OpenWorker is dramatically more affordable.

What We Didn’t Like

No Linux build yet. This is the biggest gap. OpenWorker ships for macOS 12+ (Apple Silicon, signed and notarized) and Windows 10/11 x64, but there’s no Linux build. The Hermes Agent comparison is particularly relevant here — Hermes runs on Linux. If you’re a Linux user, you’ll need to build from source (requires Python 3.10+, Node 20+, Rust toolchain), which is non-trivial.

Windows installer is not code-signed. The Windows build triggers a SmartScreen warning because the installer isn’t signed. It’s safe — it’s open-source MIT code you can inspect yourself — but the warning will spook less technical users. Microsoft’s code-signing process is expensive and slow, and for an open-source project that just launched, this is understandable but still annoying.

Beta quality — 76 open issues. OpenWorker launched two days ago. It works, but it has rough edges. We hit a few minor bugs — a connector that needed re-authentication mid-task, a scheduled automation that didn’t fire on time, some UI jank in the Tauri window. The team is actively polishing, but this is not v1.0 software yet.

No official inference service. OpenWorker requires you to bring your own API keys. For developers, this is normal. For non-technical users, it’s a barrier. You need to sign up for OpenAI, Anthropic, or another provider, get an API key, and configure it in the app. There’s no “click to start” experience.

Steep setup for non-developers. Even beyond API keys, you need to configure connectors (Slack OAuth, GitHub token, etc.), understand the permission engine, and set up scheduled automations. OpenWorker is designed for power users. If you’re not comfortable with API keys and OAuth flows, you’ll struggle.

Knowledge worker focused — not a coding agent. OpenWorker can read and write files, run shell commands, and interact with git, but it’s not designed for code editing. If you need an AI pair programmer for software engineering, use Claude Code or Codex CLI. OpenWorker is for the other 80% of knowledge work.

The Verdict

OpenWorker is the best free open-source desktop AI coworker for knowledge workers in 2026. Period. The combination of model agnosticism, local-first privacy, 25+ integrations, a sophisticated permission engine, scheduled automations, and an MIT license is unmatched. It’s not for everyone — Linux users, non-developers, and code-first users should look elsewhere — but for its target audience, it’s a breakthrough.

DOWNLOAD OPENWORKER — free, open-source, MIT licensed.

Claude Code: Best for Code Editing (But Not a Coworker)

Claude Code is Anthropic’s CLI-based coding agent. It’s excellent at what it does — editing code, running tests, debugging, and refactoring — but it’s not designed for knowledge work. It’s a terminal tool, not a desktop coworker.

What We Liked

Best-in-class code reasoning. Claude Opus 4.8 and Sonnet 4.6 are exceptional at understanding codebases, planning refactors, and writing correct code. For software engineering tasks, Claude Code is the gold standard.

MCP protocol support. Claude Code supports the Model Context Protocol for tool integration, giving it access to filesystems, databases, and external APIs through MCP servers.

Approval gates. Claude Code asks before potentially destructive operations, which is essential for a coding agent that can modify your filesystem.

What We Didn’t

CLI-only. No desktop GUI. No visual feedback. You work entirely in the terminal. For knowledge work — reading documents, managing calendars, triaging Slack — this is a non-starter.

Claude-only. You’re locked to Anthropic’s models. No switching to OpenAI, Google, DeepSeek, or local models. If Anthropic raises prices or has an outage, you’re stuck.

Remote sessions. Your data goes through Anthropic’s servers. No local-first option. For privacy-sensitive work, this is a dealbreaker.

No scheduled automations. Claude Code can’t run recurring tasks. No daily briefs, no weekly reports, no automated watches.

Pricing. Claude Code is part of Claude Pro ($20/mo) or Max ($100/mo). For heavy usage, the cost adds up quickly.

The Verdict

Use Claude Code for code editing. Don’t use it for knowledge work. It’s a specialized tool for a specific job, and it does that job well — but it’s not a general-purpose AI coworker.

TRY CLAUDE CODE — part of Claude Pro/Max.

Codex CLI: Best for OpenAI-Native Code Editing

Codex CLI is OpenAI’s CLI-based coding agent, built on GPT-5.6 Sol. Like Claude Code, it’s designed for code editing in a sandboxed environment.

What We Liked

GPT-5.6 Sol integration. OpenAI’s latest model is powerful, fast, and capable of complex code reasoning. For Python, TypeScript, and React code, Codex CLI produces excellent results.

Sandboxed execution. Codex CLI runs code in an isolated sandbox, preventing accidental damage to your system. This is safer than Claude Code’s filesystem-level access.

Apache 2.0 license. Codex CLI is open source under Apache 2.0, which is more permissive for commercial use than many alternatives.

What We Didn’t

CLI-only. Same limitation as Claude Code. No desktop GUI, no visual workspace, no integration with productivity tools.

OpenAI-only. You’re locked to OpenAI’s models. No switching to Anthropic, Google, DeepSeek, or local models.

Cloud sandbox. Codex CLI runs in OpenAI’s cloud sandbox, not locally. Your code and data go through OpenAI’s servers. No offline capability.

No tool integrations. Codex CLI has filesystem and sandbox access but lacks the 25+ pre-built integrations that OpenWorker offers. No Slack, no GitHub issues, no HubSpot, no Notion.

No scheduled automations. Like Claude Code, Codex CLI can’t run recurring tasks.

The Verdict

Codex CLI is a solid coding agent for OpenAI users. But it’s even more limited than Claude Code for knowledge work — no integrations, no scheduled tasks, and cloud-only execution make it unsuitable as a general-purpose AI coworker.

TRY CODEX CLI — part of ChatGPT Plus.

Hermes Agent: Closest Open-Source Competitor

Hermes Agent is an open-source (MIT) full-system agent that runs locally and supports multi-model execution. It’s the closest competitor to OpenWorker in the open-source space.

What We Liked

Multi-model support. Like OpenWorker, Hermes supports multiple model providers. You’re not locked in.

Full local execution. Hermes runs entirely on your machine. No cloud dependency, no data leaving your system.

Cron-based automations. Hermes supports scheduled task execution via cron, which is rare in this space.

Whisper speech input. Hermes includes speech-to-text via OpenAI Whisper, similar to OpenWorker’s Rust STT sidecar.

What We Didn’t

No desktop GUI. Hermes is CLI and messaging-based. No Tauri desktop app, no React UI, no visual workspace. OpenWorker’s desktop GUI is a significant UX advantage.

Fewer pre-built integrations (15+ vs 25+). Hermes has solid MCP support but fewer ready-to-use connectors. You’ll spend more time configuring integrations.

Less polished permission engine. Hermes has approval gates, but OpenWorker’s four risk classes and five permission modes are more sophisticated and better documented.

Smaller community and ecosystem. OpenWorker launched with 2,651 stars and backing from Andrew Ng. Hermes has a smaller community, fewer contributors, and less active development.

The Verdict

Hermes Agent is a capable open-source alternative, especially for Linux users who can’t run OpenWorker yet. But OpenWorker’s desktop GUI, larger integration library, more sophisticated permission engine, and stronger community backing make it the better choice for most users.

Pricing Breakdown

Here’s the honest cost picture for each tool:

Tool	Software Cost	Model Cost (per 1M input tokens)	Total Monthly (typical use)
OpenWorker + Ollama	$0 (MIT license)	$0 (local)	$0
OpenWorker + DeepSeek V4	$0 (MIT license)	~$0.04	~$2-5
OpenWorker + Claude Fable 5	$0 (MIT license)	~$0.12	~$10-20
OpenWorker + GPT-5.6 Sol	$0 (MIT license)	~$0.15	~$15-30
Claude Code	$20-100/mo (Pro/Max)	Included	$20-100
Codex CLI	$20/mo (ChatGPT Plus)	Included	$20
Hermes Agent + Ollama	$0 (MIT license)	$0 (local)	$0

The math is clear. OpenWorker with Ollama costs $0. OpenWorker with DeepSeek V4 costs a few dollars a month. Even with premium models like Claude Fable 5 or GPT-5.6 Sol, you’re paying $10-30/month — less than Claude Pro alone, and you get a full desktop AI coworker instead of a CLI coding agent.

If you want privacy and zero cost, use Ollama for local inference. If you want the best reasoning for complex tasks, use Anthropic’s API with Claude Fable 5. If you want the best value, use OpenAI’s API with GPT-5.6 Sol or DeepSeek V4.

Head-to-Head: OpenWorker vs Claude Code — Which Should You Buy?

This is the most common comparison, so let’s be direct.

Choose OpenWorker if:

• You need a desktop AI coworker for knowledge work (research, reports, Slack triage, CRM updates, calendar management)
• You want model flexibility — switch between OpenAI, Anthropic, Google, DeepSeek, or local models
• Privacy matters — you want all data to stay on your machine
• You need scheduled automations for recurring work
• You want 25+ pre-built integrations plus MCP extensibility
• You want to pay $0 for the software and only for the model API keys you use

Choose Claude Code if:

• You primarily need AI-assisted code editing, debugging, and refactoring
• You’re already in the Anthropic ecosystem and happy with Claude models
• You prefer a CLI workflow and don’t need a desktop GUI
• You don’t need scheduled automations or extensive tool integrations
• You’re willing to pay $20-100/month for the privilege

The honest truth: most knowledge workers should get OpenWorker. It’s free, it’s more capable for non-code tasks, and it gives you model flexibility. Keep Claude Code in your back pocket for heavy coding sessions, but make OpenWorker your daily driver.

Bottom Line Final

So which one should you buy?

If you’re a developer, ops engineer, or power user who wants a privacy-preserving, model-agnostic AI coworker that delivers finished work instead of chat, get OpenWorker. It’s free (MIT license), runs locally, supports 30+ models, integrates with 25+ tools plus MCP, and has a sophisticated permission engine that keeps you in control. No other tool in this comparison combines all of these features.

If you primarily need AI code editing, get Claude Code or Codex CLI — they’re better at software engineering. If you’re on Linux, consider Hermes Agent until OpenWorker ships a Linux build.

But for the vast majority of knowledge workers — people who research, write, manage calendars, triage Slack, update CRMs, and generate reports — OpenWorker is the winner. It’s the first desktop AI tool that actually delivers finished work, not just advice.

DOWNLOAD OPENWORKER — free, open-source, MIT licensed. Run it with Ollama for free local inference, or bring your own API key from OpenAI or Anthropic.

FAQ

Is OpenWorker really free?

Yes. OpenWorker is MIT-licensed open-source software. You pay $0 for the application. The only costs are the AI model API keys you choose to use. Run it with Ollama locally and the total cost is zero — just your hardware. Use DeepSeek V4 and it costs pennies. Use Claude Fable 5 or GPT-5.6 Sol and it costs $10-30/month depending on usage.

How is OpenWorker different from Claude Code or Codex CLI?

OpenWorker is designed for knowledge work — research, reports, Slack triage, CRM updates, calendar management, document creation. Claude Code and Codex CLI are designed for code editing. OpenWorker has a desktop GUI (Tauri 2 + React), supports 30+ models, keeps data local, has 25+ pre-built integrations, and supports scheduled automations. Claude Code and Codex CLI are CLI-only, locked to a single model provider, and have no scheduled automations.

Does OpenWorker work on Linux?

Not yet. OpenWorker ships for macOS 12+ (Apple Silicon, signed and notarized) and Windows 10/11 x64 (not code-signed). There is no Linux build. You can build from source if you have Python 3.10+, Node 20+, and the Rust toolchain, but it’s not a trivial process.

Do I need an API key to use OpenWorker?

Yes, unless you use Ollama for local inference. OpenWorker does not have its own inference service. You bring your own API key from OpenAI, Anthropic, Google, or any of the 30+ supported providers. For local inference, set up Ollama and OpenWorker will use it with no API key required.

Is my data private with OpenWorker?

Yes. OpenWorker is local-first. Conversations, tokens, and API keys stay on your machine. Model calls go directly from your computer to the provider — there is no OpenWorker cloud intermediary. The only optional cloud component is an Auth0 PKCE OAuth broker for one-click connector setup, and tokens never get stored in the cloud. OpenWorker works fully signed out with manually pasted credentials.

Disclosure: Some links in this post are affiliate links. If you purchase through these links, we may earn a small commission at no extra cost to you. OpenWorker itself is free and open-source — we only earn commissions on Ollama, OpenAI API, and Anthropic API referrals if you choose to use those services.