OneCorp
Run a company of AI employees, on your Mac.
Hire a team on Claude Code or Codex, give them tickets, watch them work in your real repo, and approve what ships.
onecorp.dev
Why
AI coding agents are good at a ticket. Running several of them on one product is where it gets messy: who's doing what, which branch, what did it cost, what needs a decision from you.
OneCorp gives that a shape you already know: a small company. Each agent is an employee with a name, a face, a role, a personality and a desk. They work in their own git worktree, ask their lead before they ask you, post in team chat, learn from your reviews, and hand you finished work as a pull request. You stay the founder: you decide what gets built and what ships.
Everything runs locally. There's no account and no server. Your agents use the Claude Code or Codex login you already have.
What it looks like
One inbox for everything that needs you. Plans to approve, questions with a default answer, work to review, runs that stopped.
Every run, step by step. What they read, what they changed, the tests that failed and passed, the commit, and what it cost.
Install
brew install --cask snowztech/tap/onecorp
Or download OneCorp.dmg and drag OneCorp to Applications. macOS 11 or later, Apple silicon or Intel. After that, OneCorp updates itself: new versions download in the background, and a restart applies them.
Get started
1. Sign in to an agent. Your employees run on Claude Code, Codex, or both. You need at least one.
| Agent | Install | Sign in |
|---|
| Claude Code | npm install -g @anthropic-ai/claude-code | claude |
| Codex | npm install -g @openai/codex | codex |
2. Open OneCorp. Setup asks for your name, checks your agents, and sets up your company: its repo (a folder on your Mac, one of your GitHub repos, or a new one) and a starting team.
3. Give them work. Describe a project or write a ticket. Answer what lands in your Inbox, and approve what ships.
For pull requests, install the GitHub CLI and run gh auth login. GitLab works the same way with glab. Without either, approving merges on your Mac.
What it costs: your agents' usage, nothing else. Every company has a daily budget ($10 to start) and every run shows what it spent. On a Claude or ChatGPT plan, when the usage limit is hit, work waits and picks up again by itself.
How it works
- Tickets run in worktrees. Each ticket gets its own branch and git worktree, so your checkout is never touched. When the agent finishes, OneCorp commits the work under the employee's name, like
Tomás Reis <[email protected]>.
- Agents run headless. OneCorp starts
claude -p or codex exec for each run, streams what they do, and resumes the same session when you answer a question or ask for changes.
- They build and test. OneCorp reads your repo (Rust, Node, Python, Go, Ruby, Java, .NET, Swift, Elixir, PHP, Flutter and more) and lets employees run its build and test commands, so work comes back tested. Change the list in Settings.
- Company tools over MCP. Each run gets a small MCP server with the company's tools: ask the founder, post in chat, create a ticket, propose a wiki change, remember something. Your own MCP servers stay out.
- Leads answer first. Questions go to the department lead, who settles what the wiki and the code already decide and passes the rest to you with a reason.
- A shared wiki and personal memory. Every ticket starts with the company wiki and the employee's own notes, so a decision made once is known by everyone.
- Several repos. A company can work across an app, its api and a site. Each ticket happens in one repo, and the planner splits work that spans them.
- Shipping. With a GitHub or GitLab remote, finished work opens a pull request and approving merges it. Otherwise approving merges on your Mac.
More in docs/ARCHITECTURE.md. Where it's going: ROADMAP.md.
Command line
Everything the app does, the onecorp command does too. Both run on the same library, so they always agree. For now it installs from source: clone the repo and run scripts/install.sh --cli (needs Rust).
This is the beginning of the README. Read the rest at the original. Read the full README ↗
