JVCompany Star on GitHub
Open source · Self-hosted · MIT

Your software company, run by AI agents.

JVCompany is a free multi-agent platform that turns Claude Code, OpenAI Codex, Gemini CLI and any LLM into a real team: hand one request to the CEO, and ten AI departments plan, design, code, test, review and ship it — live, inside a 3D office building.

Get started free View source on GitHub
  • Claude Code
  • OpenAI Codex
  • Gemini CLI
  • Cursor
  • DeepSeek
  • Ollama
JVCompany dashboard showing running projects, tasks, approvals and the 5-floor 3D company building
3D office floor in dark mode with AI agents working at their desks

Works with the AI you already use

  • Claude Code
  • OpenAI Codex
  • Gemini CLI
  • Cursor Agent
  • Copilot CLI
  • Qwen Code
  • OpenCode
  • Claude API
  • OpenAI
  • Gemini
  • Grok
  • DeepSeek
  • OpenRouter
  • Groq
  • Mistral
  • Ollama
  • LM Studio
Features

Everything an AI software team needs

An org chart, a workflow, permissions, a budget and a boss — so your coding agents deliver finished, reviewed work instead of one-off answers.

10 AI departments

CEO, BA, PM, Scrum Master, UI/UX Studio, Architect/DevOps, Senior Dev, QA, Tester and Bug Fixer — each with a role-specific prompt, permissions and model.

Delegate to the CEO

One request in. The director scopes it, builds a hand-off chain, grants per-task permissions and writes the final report.

Any model, any plan

Subscription CLIs, API keys, OpenAI-compatible endpoints, local models — or ChatGPT/Claude.ai in the browser. Mix models per department.

Live 3D office

Five floors, five projects in parallel. Agents code at their desks, meet, wait for approvals and play billiards — driven by real task state.

Permissions & approvals

Read-only, read & write, auto in project or full computer control. Anything above the level becomes an approval request.

Rate-limit aware

Detects usage limits, waits for the exact reset and resumes — or switches to a fallback connection. Claude Code resumes the same session.

Costs & budgets

Tokens and cost for every run; budgets per company, department or project with automatic pause at 100 %.

Goals & routines

Goal trees, JV-123 task IDs, multi-step approvals, and routines triggered by cron, webhook (HMAC) or API.

Secrets, skills & MCP

AES-256-GCM secret vault, SKILL.md skills, plugins, MCP connections, agent evals and a full activity log.

How it works

One request. A whole team on it.

The director turns your request into a hand-off chain. Each step starts automatically with the previous step's output; failures go back to the director, and you approve anything risky.

  1. AskChat with the CEO, pick a project folder, paste screenshots.
  2. AssignThe tower CEO finds a free floor — or queues the project.
  3. PlanThe floor director scopes the work and builds the chain with per-step permissions.
  4. BuildDepartments work in your folder; results flow to the next step automatically.
  5. ReviewQA approves, the director reports, and you sign off on anything risky.
3D office

Watch your AI company at work

A five-floor building rendered with three.js. Every character reflects a real agent and a real task.

  • Typing at desks while a department is running.
  • Meeting room when the floor director hands out new work.
  • Standing by when an agent needs your approval.
  • Billiards & coffee when they're idle.
  • You're a CEO too — ride the elevator between floors and give orders on the spot.
3D office floor with meeting room, desks, billiards lounge and a chat with the AI CEO
AI connections

Use the subscriptions you already pay for

Each department picks its own connection, a model tier (Premium / Balanced / Fast) and a fallback.

  • Subscription CLIs — Claude Code, Codex, Gemini CLI, Cursor, Copilot CLI, Qwen Code, OpenCode and custom commands, auto-detected.
  • API keys — Claude, OpenAI, Gemini, Grok, DeepSeek, OpenRouter, Groq, Mistral and any OpenAI-compatible endpoint.
  • Local models — Ollama and LM Studio, nothing leaves your machine.
  • External agents — Hermes, OpenClaw, Cursor Cloud Agent or any HTTP webhook.
AI connections page with a DeepSeek API key and a Claude Code subscription showing usage limits and token costs
Screenshots

A tour of the office

English UI shown — the app ships in 10 languages, English by default.

Dashboard with live projects and the company tower
Dashboard — projects, tasks, approvals at a glance.
Starting a new chat with the CEO on a free floor
New chat — every conversation becomes a project.
Project page with stages, conversation, tasks, documents and files
Project — stages, chat, tasks, docs, files.
Staff roster of AI departments
Staff — one roster for all floors.
Task list across all departments
Tasks — every department, every project.
Approval requests from AI agents
Approvals — allow once, per task, or deny.
Org chart of AI agents
Org chart — heartbeats, budgets, history.
Costs and budgets dashboard
Costs — spend per run and budget.
Routines scheduled by cron or webhook
Routines — cron, webhook or API.
Company goal tree
Goals — company to task.
Shared skills for departments
Skills — SKILL.md for every team.
Agent evaluation suites and runs
Evals — measure your agents.
Quick start

Up and running with one command

You need Node.js 20+ and MongoDB (local, Docker or Atlas). Agent CLIs you have signed in to are detected automatically.

  • create-jvcompany downloads the source, writes .env.local, checks MongoDB and installs everything.
  • A 4-step setup creates your company, team and first project.
  • Prefer a desktop app? npm run desktop (Windows, macOS, Linux).
  • Listens on 127.0.0.1 only — your office stays private.
npx create-jvcompany@latest   # → http://127.0.0.1:2501

Or install manually:

git clone https://github.com/inuxmax/JVcompany.git
cd JVcompany
npm install
cp .env.example .env.local   # set MONGODB_URI
npm run dev                  # http://127.0.0.1:2501
Safety

You decide what agents may touch

Four permission levels per project and department. Anything above the level is turned into an approval request.

LevelReadWriteShellOutside projectMouse / keyboard
Read-only✅❌❌ask❌
Read & write✅✅askaskask*
Auto inside project✅✅✅askask*
Full computer access✅✅✅✅✅*

* Computer control must also be enabled in Settings.

FAQ

Frequently asked questions

Is JVCompany free?

Yes. It is open source under the MIT license and runs on your own computer. You only pay for the AI you connect — or nothing with local models (Ollama, LM Studio) or subscriptions you already have.

Which AI models and coding agents are supported?

Claude Code, OpenAI Codex, Gemini CLI, Cursor Agent, GitHub Copilot CLI, Qwen Code, OpenCode and custom CLIs; API keys for Claude, OpenAI, Gemini, Grok, DeepSeek, OpenRouter, Groq, Mistral and any OpenAI-compatible endpoint; local models via Ollama and LM Studio.

Does my code leave my machine?

The app runs locally and binds to 127.0.0.1. Code is only sent to the AI provider you choose for a department — use a local model if nothing may leave your computer.

Can AI agents break my computer?

Every project and department has a permission level from read-only to full access. Anything above it becomes an approval request you must accept, and computer control is off until you enable it.

What happens when I hit a usage limit?

The run pauses, the reset time is read from the error, and the task resumes automatically — or switches to the department's fallback connection.

Hire your first AI team today

Free, open source and private by default. If you like it, a ⭐ on GitHub helps others find it.

⭐ Star on GitHub Install now