Skip to content
Get packs

Guide · Open source

How to use OpenWorker: install, connect a model and run tasks

Updated · written with AI from the project's own documentation and checked against our catalogue data

OpenWorker is an open-source desktop app that runs an AI agent on your own machine. Instead of only chatting, it works on tasks such as reviewing code for security issues, drafting documents, or sorting an inbox. This guide walks through downloading the app, connecting a model, and running your first task with approvals.

OpenWorkerandrewyng/openworker · ★ 18.9k · MIT

Quick start

  1. 1Check requirements. Confirm you have macOS 12+ on Apple Silicon or Windows 10/11 on x64, and either a model API key or Ollama installed.
  2. 2Download the app. Use the macOS or Windows download link from the README or openworker.com. On Windows, expect a SmartScreen warning because the build is not yet code-signed.
  3. 3Open the app. Launch OpenWorker after installing it on your machine.
  4. 4Add a model. Go to Settings → Models & Keys, paste your provider key, or point the app at a local Ollama server.
  5. 5Describe a real task. Ask for a specific outcome, such as a customer brief or a security review of a repository.
  6. 6Approve actions. When the agent asks before sending, changing or running something, approve it or redirect the work.
  7. 7Review the result. Open the finished file or deliverable and check the audit trail for what ran and who approved it.

What you need before you install OpenWorker

The desktop app has platform requirements, and you need a way to reach a language model. You can use an API key from a model provider, or run local models through Ollama.

  • macOS 12 or later on Apple Silicon, or Windows 10/11 on x64
  • A model API key from a supported provider, or Ollama installed for local models
  • An MIT-licensed, open-source project, so you can read the code

Download and install the desktop app

Use the download links on the OpenWorker README or at openworker.com. The macOS build is signed and notarized and updates itself. The Windows build is not yet code-signed, so Windows SmartScreen will show a warning during install. The README says signing is in progress.

The README also includes a git clone command for the source repository if you want to read or build the code yourself.

  • macOS: signed and notarized, with auto-updates
  • Windows: expect a SmartScreen warning until code signing is finished
git clone https://github.com/andrewyng/openworker

Add a model key or connect Ollama

Open the app and add a model key, or point it at Ollama. You choose the provider, paste your key, and can switch later. Each provider card keeps its own models, so disconnecting one does not fall back to another key.

For OpenRouter, Settings → Models & Keys offers two routes. The OpenRouter (API keys) option takes a key you paste. The OpenRouter account (Subscriptions) option signs you in through the browser and uses your OpenRouter credits. On a remote backend, choose Use a manual code and paste the code OpenRouter shows.

  • Cloud providers include OpenAI, Anthropic, Google Gemini, Mistral, DeepSeek and others
  • Local options include Ollama, llama.cpp, vLLM, LM Studio, Jan, Docker Model Runner and mlx-lm
  • A curated list shows models verified for tool-calling. Other model strings work at your own risk

Run your first task and approve actions

Describe the outcome you want, such as preparing a customer brief or drafting a report. OpenWorker breaks the task into steps and works across your files, terminal and connected apps. Before anything consequential, such as sending a message, changing a calendar or running a command, it asks you to approve or redirect.

Actions are approval-gated by default. Auto-approve mode exists, but a reviewer model escalates anything it is unsure about to you.

  • Describe the outcome, not a list of steps
  • Approve, redirect or deny when the agent checks in
  • Review the audit trail to see who approved each tool call and why

Use security review and connected tools

The Security review coworker scans a codebase and its dependencies. Findings come from deterministic scanners such as semgrep plus model reasoning. Proposed fixes are re-scanned and diff-reviewed before you approve them.

OpenWorker lists 25+ integrations, including GitHub, Slack, Jira, Notion, Gmail and Google Calendar. Any tool reachable over MCP can also be added, with per-tool control. Scheduled automations, such as a morning brief or a weekly report, land in the app with full transcripts.

  • Security review: scanners plus model reasoning, with re-scan and diff review
  • Connectors: GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, monday.com, Gmail, Google Calendar
  • Automations: recurring reports and channel watches on a schedule

Understand the sandbox and governance options

Governance is built into the architecture. Dangerous or irreversible operations are human-only, and no mode, including full auto-approve, lowers these floors. Unattended runs never self-approve; their requests wait in an inbox until a person answers.

Commands and file tools can run inside an NVIDIA OpenShell sandbox. On a Mac, the built-in macOS sandbox gives similar protection with nothing to install. On Windows, a hidden local account is used. The README describes reviewer verdicts as judgments, not guarantees.

  • Hard floors: dangerous operations always escalate to you
  • Audit trail: every tool call is recorded with its approval source
  • Sandbox: OpenShell, macOS built-in sandbox, or a Windows hidden account

What people use it for

  • Scanning a codebase and its dependencies for security risks, with fixes re-scanned before you approve them
  • Preparing a customer call by pulling context from your CRM and inbox
  • Drafting a weekly report or morning brief on a schedule
  • Triaging an inbox and handling calendar changes with approval
  • Working an incident by gathering context across tools and drafting the timeline

Frequently asked questions

Is OpenWorker free?

The README names the project as open source under the MIT licence. Model providers charge for API usage separately, unless you run local models through Ollama.

What are the requirements for OpenWorker?

The desktop app needs macOS 12 or later on Apple Silicon, or Windows 10/11 on x64. You also need a model API key or Ollama for local models.

Does OpenWorker work with Ollama?

Yes. The README says you can run fully local models through Ollama, and the app can also connect to llama.cpp, vLLM, LM Studio, Jan, Docker Model Runner and mlx-lm.

Why does Windows show a SmartScreen warning?

The README says the Windows build is not yet code-signed, so SmartScreen will warn. The README says signing is in progress.

Is OpenWorker stable enough for daily use?

The README describes OpenWorker as in open beta. It is described as fully usable, but with rough edges that the team is still polishing, and issues are welcome on GitHub.

Are there alternatives to OpenWorker?

This analysis lists no comparable projects. You could compare the feature list in the README with other AI agent tools before choosing.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

OpenWorker suits developers and technical users who want an AI agent on their own machine, with their own model keys, approval gates, and optional local models through Ollama. It is a poor fit if you need a polished, stable product or a warning-free Windows install, or if you want an agent to act unsupervised on consequential tasks.

Install options & downloads