OpenCompany
An open-source OS for self-improving AI employees that run on your own computer and turn tokens into work.
- Source repo
- zeenie-ai/OpenCompany
- Stars
- ★ 970
- Last updated
- 1d ago
- License
- MIT
- Primary language
- Python
- FA score
- 44/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A solo developer or small team hiring a Builder: it writes and runs code, keeps dev servers alive, opens pull requests, deploys, and manages cloud resources via GitHub, Vercel, Cloudflare and Google Cloud.
- Not a fit if
- Teams that need an iOS mobile workspace (only Android emulator on Windows x64 is supported)
- Teams unwilling to store API keys locally or to run self-hosted infrastructure
- Individuals wanting a ready-made hosted chatbot with no connector setup
- Source review
- 44/100 · Major gaps
What does this agent do, and when should you use it?
OpenCompany is an MIT-licensed operating system for AI employees that you install on Windows, macOS or Linux, or self-host with the install script, Docker Compose or from source. An employee is not one chatbot but a small team: a lead that understands the job, dispatches work, checks results and reports back, plus specialist agents wired together on a drag-and-drop canvas with no code. Employees connect to Gmail, Google Calendar, Drive, Sheets, Tasks, Contacts, Microsoft 365, WhatsApp, WhatsApp Business, Telegram, Discord, X, Stripe, GitHub, Vercel, Cloudflare, Google Cloud, an Android emulator, a browser, search, scrapers and a knowledge base — 140+ tools in total — and wake on new mail, customer messages or a schedule. Models come from OpenAI, Anthropic, Google, xAI, DeepSeek, Kimi, Mistral, Groq, Cerebras, Sarvam and OpenRouter, or run locally for free through Ollama, LM Studio or any OpenAI-compatible server. Improvement is never model retraining: it lives in editable memory, notes and plain-text skills (78 shipped), plus a live browser workspace and a Windows-x64 Android workspace.
The flow is Hire → Build the team → Start → Review. On Home you describe a job in plain words, OpenCompany drafts the employee's setup (which apps it uses, when it works, what it checks with you first), you adjust and press Hire; alternatively you switch to Dev mode and drop an AI Employee onto the canvas, connect specialist agents and give each the tools for its part of the job (email, browser, code, messaging, payments). Teams run in the background and are event-driven: a new email, a customer message or a scheduled time wakes them, the lead reviews every result from the team and sends it back if it is not right, and you can require approval before anything goes out. Output lands in the employee's Workspace, where you read what it produced and edit what it remembers; skills are short plain-text playbooks that take effect on the next turn and are managed under Settings > Skills. The runtime also ships a live browser workspace — Home and Dev mode can watch an employee browse, let you take control to sign in, then hand it back, all inside a dedicated OpenCompany profile of your installed Chrome, Edge or Chromium — plus a local, controllable Android emulator workspace on Windows x64, and speech and translation so employees can listen, talk and work in other languages.
- A solo developer or small team hiring a Builder: it writes and runs code, keeps dev servers alive, opens pull requests, deploys, and manages cloud resources via GitHub, Vercel, Cloudflare and Google Cloud.
- A community or content operator needing a Grower: it publishes to X, WhatsApp channels, Telegram and Discord on a schedule, while an analyst does social-media data and web research on the market.
- Someone with a heavy inbox running a Chief of Staff on Gmail or Microsoft 365: it reads and answers mail, keeps the calendar, orders files, sheets, tasks and contacts, and delivers a morning summary.
- A customer-facing team using a Front Desk that replies on WhatsApp, WhatsApp Business, Telegram and Discord in the customer's language (voice optional) and escalates what it cannot resolve.
- Research work handled by a Researcher: it browses, searches, scrapes, reads documents and images, and files findings into a knowledge base the whole company can query.
- A payments workflow run by a Treasurer: it operates through Stripe and reacts the moment a payment event fires.
How do you install or deploy this agent?
Desktop app (recommended): download the package for your machine from releases/latest — the Windows x64 .exe, the macOS -arm64.dmg or -x64.dmg, or the Linux .AppImage / .deb. On first launch the app sets itself up (a one-time download of a minute or two). Builds are not code-signed yet, so macOS requires allowing the app under System Settings > Privacy & Security and Windows SmartScreen needs "More info > Run anyway".
Terminal install for servers and headless machines:
curl -fsSL https://opencompany.sh/install.sh | bash # macOS / Linux
iwr -useb https://opencompany.sh/install.ps1 | iex # Windows PowerShell
company startThe script installs bun, Python and uv when missing, then the @zeenie-ai/opencompany package; data lives in ~/.opencompany, shared with the desktop app.
Docker self-hosting:
git clone https://github.com/zeenie-ai/OpenCompany.git
cd OpenCompany
docker compose up -d --buildThat builds the image from source and starts one container with data in the opencompany-selfhost_data volume; open http://localhost:5678 and register the owner account (the port is published on this machine only). Running from source needs bun 1.4+ and Python 3.12:
git clone https://github.com/zeenie-ai/OpenCompany.git
cd OpenCompany
bun run build
bun run devHow do you use this agent?
On first launch the app opens Home; connect an AI provider there or under Settings > Connectors, then describe a job in plain words to hire your first employee, or start from a ready-made bundle in Settings > Plugins. Switch to Dev mode for the workflow editor, where three example employees sit in the workflow sidebar and can be opened to see how they are assembled; drop an AI Employee onto the canvas, connect specialists and configure each one's tools. Watch running teams from their card on Home or the canvas, open the Workspace to see what they made, and change behaviour by editing memory or the plain-text skills; on the Talk page you can ask an employee to take on something new and it can add a tool or learn a skill within the rules you hired it with, taking effect in that conversation immediately and across all its work once you press Apply. Hand control of the browser workspace back and forth when a sign-in is needed. For shared or cloud use, add a login; one command deploys to Google Cloud.
What are this agent's strengths and limitations?
- The team structure is the product, not a prompt: a lead understands the job, hands out work, reviews every result and sends it back if it is not right before it reaches you.
- Model-provider neutral by design — OpenAI, Anthropic, Google, xAI, DeepSeek, Kimi, Mistral, Groq, Cerebras, Sarvam and OpenRouter, plus free local inference via Ollama, LM Studio or any OpenAI-compatible server such as llama.cpp or vLLM.
- Fully inspectable learning: memory, notes, skills and canvas pieces are readable and editable, and the project states it never retrains a model, so there is no black box.
- Improvement happens at runtime: a tool added on the Talk page works in that conversation right away and, with Apply, becomes part of all the employee's work without redeploying.
- MIT-licensed with three documented delivery paths — signed-less desktop installers, the install script, and Docker Compose — and desktop and terminal installs share the ~/.opencompany data directory.
- Desktop builds are not code-signed yet, so macOS and Windows SmartScreen require manual approval — friction for fleet distribution and compliance review.
- Connector and provider presets are deliberately narrow; the README says a new first-class preset needs a reason beyond "my service could be in the dropdown too", so niche SaaS has to arrive via the Apify
customoption, TikHubcall, or a self-saved OpenAI-compatible endpoint. - The mobile workspace is limited to a local Android emulator on Windows x64 and requires accepting the SDK license; iOS is explicitly not supported yet.
- The software is free but real use generally depends on paid model API keys or local inference hardware, and using the 140+ tools means supplying accounts and credentials for each service.
- There is no vendor-hosted product as the main path: self-hosting means Docker or the script and operating it yourself, and running from source adds two runtime prerequisites, bun 1.4+ and Python 3.12.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenCompany This agent | 44 · Major gaps | Desktop appFree + model costs | ★ 970 | 1d ago | Python | OpenAI API · Claude API |
| BagIdea Office | 71 · Some gaps | Desktop appFree + model costs | ★ 242 | 1d ago | JavaScript | Claude Code · OpenAI API |
| Atom Platform | 59 · Major gaps | CLIFree + model costs | ★ 901 | 5d ago | Python | OpenAI API · Claude API |
| GAIA — Personal AI Assistant | 73 · Some gaps | Hosted serviceFreemium | ★ 302 | 5d ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
README claims employees ask before sending or spending and that API keys are stored encrypted locally, giving partial evidence for user_confirmation (2). least_privilege scores 1: the repo advertises 140+ tools, a browser workspace, an Android emulator, Stripe payments, and cloud deployment, yet provides no permission manifest, least-privilege configuration, or sandboxing description. data_flow_transparency and sensitive_data_handling each score 1: only a single sentence about encrypted key storage, with no data-flow diagram, retention policy, or sensitive-data classification. dependency_security scores 2: package.json overrides pin minimum safe versions for multiple known-vulnerable packages (dompurify, esbuild, hono, ws), a verifiable mitigation. external_effects and rollback each score 1: README mentions pause/resume and missed-schedule catch-up, but no mechanism to roll back sent emails, published content, or executed payments. source_attribution scores 1: LICENSE credits both MachinaOs and OpenCompany contributors, package.json author is a personal email, and publisher identity is unverified by FollowAgents, leaving the attribution chain unclear.
self_consistency scores 2: package.json and pyproject.toml agree on version 0.2.1, CLI entry points company/machina match documentation, and CI workflows reference real paths. dependency_availability scores 2: dependencies are all publicly obtainable (typer, rich, anyio, psutil, platformdirs, pywin32), with platformdirs noted as a soft dependency and pywin32 as Windows-only, and fallback paths documented. failure_messages scores only 1: cli/tests/test_backend_shutdown.py shows a ValueError for invalid TEMPORAL_GRACEFUL_SHUTDOWN_SECONDS, but the repo offers no user-facing error-message conventions or troubleshooting docs, and README does not describe common failure modes.
audience_and_scenarios scores 2: README clearly separates four audiences—desktop users, server/headless users (terminal install), self-hosters (Docker), and contributors (source build)—with concrete scenarios. capability_boundaries scores only 1: it lists employee roles (Builder, Grower, Chief of Staff) but does not state model capability limits, failure scenarios, or what the agent cannot do. trigger_precision scores 1: only vague triggers are mentioned (new email, customer message, scheduled time), with no trigger syntax or priority rules. environment_fit scores 2: covers Windows/macOS/Linux, Docker, and local models (Ollama, LM Studio, vLLM), and notes Android is Windows x64 only and iOS unsupported, giving fairly complete environment guidance.
information_architecture scores 2: README is well structured with Quick Start, How It Works, The Employees, What Is in the Box, and For Developers sections, linking CONTRIBUTING.md, SETUP.md, and SCRIPTS.md. install_notes scores 2: four install paths (desktop installers, one-line terminal script, Docker, source) with a note that unsigned builds require manual macOS/Windows approval. naming_stability scores only 1: package name @zeenie-ai/opencompany matches the repo name, but the CLI keeps both company and the deprecated machina alias, and LICENSE shows the legacy MachinaOs name, indicating an incomplete rename. examples_and_faq scores 1: only video demos and three diagrams, no textual examples or FAQ. known_limitations scores 1: only unsigned builds, no iOS, and Android limited to Windows x64 are mentioned; no systematic limitations section. license scores 2: full MIT text with clear copyright years and holders. versioning_changelog scores 1: package.json and pyproject.toml agree at 0.2.1, but no CHANGELOG file exists. maintenance_responsibility scores 1: README points to Discord and GitHub Issues, but publisher identity is unverified and there is no maintainer list or response commitment.
output_usability scores only 1: README describes employee outputs (email replies, PRs, published content, payment operations) but gives no output formats, templates, or quality standards. marginal_value scores 1: its differentiation is role-based 'AI employees' and canvas orchestration, but there is no comparison with existing tools (n8n, LangGraph) or quantified benefit evidence. cost_benefit scores 1: it claims bring-your-own-keys, no subscription, no usage limits, and free local models, but provides no token consumption, cost estimates, or value comparison against alternatives.
claim_traceability scores only 1: specific numbers such as '140+ tools', '78 skills', and '12 themes' have no corresponding manifest or file index to check against. cross_source_corroboration scores 1: package.json, pyproject.toml, CI workflows, and README broadly agree on version, entry names, and install methods, but core functional claims (self-improvement, memory, skills) lack code or test corroboration. fact_inference_separation scores 1: README mixes marketing language ('turning LLM tokens into work and dollars', 'gets better the longer it works') with factual description, without separating verified facts from vision.
- Publisher identity is unverified by FollowAgents, and LICENSE credits both MachinaOs and OpenCompany contributors, leaving the attribution chain unclear; verify the maintaining entity before adoption.
- The repo advertises 140+ tools, a browser workspace, an Android emulator, Stripe payments, and cloud deployment, but provides no permission manifest or least-privilege configuration, so actual privilege scope cannot be confirmed from static files.
- Specific numbers such as '140+ tools', '78 skills', and '12 themes' have no corresponding manifest or index and cannot be verified in a static review.
- Install scripts use curl | bash and iwr | iex, and builds are unsigned, requiring manual bypass of macOS/Windows security prompts, which carries supply-chain and trust risk.
- No CHANGELOG, FAQ, or systematic known-limitations section exists, leaving long-term maintenance and upgrade paths unclear.