A virtual office for your team of AI agents
When you run more agents than you can watch, you lose track of time, cost and results.
Arc keeps every project, task and hour in one place, for people and Claude alike. Your usual Claude Code and your own subscription, nothing new to learn.
The agents worked all week. How many hours went to which client, nobody can say.
Every session spends tokens. Which agent and which project they went to is a guess.
What got done and where it got stuck lives only in your head, until you forget it.
Every hour is counted. Arc measures time from Claude's own sessions and ties it to the task. You see who worked on what, people and agents alike.
Tasks, priorities and a roadmap. Every session starts from a task, and every commit points back to it.
Developer, analyst, designer: each agent knows its project and its rules, and messages you in Telegram when it is done.
The strong model for hard work, a fast one for routine, and the tokens every task took in view.
Turn the chaos into a system you can run.
A project per client or product. Connect its repository or start empty. Tasks, wiki and notes live with it.
Developer, analyst, designer. Each one gets the project's context, its own skills and rules, and the model that fits the work.
Start Claude with arc <project>, pick a task and work. Hours and commits land on the task by themselves, and Telegram tells you when an agent is done.
Pain: Three clients, one Claude subscription. At the end of the month you rebuild the hours from memory.
Outcome: Hours per client and per task are already counted, and the report for the client is one click away.
Pain: People and agents work in parallel. Who is doing what, you find out at the stand-up.
Outcome: One task list for people and agents, with priorities, owners and the hours each task took.
Pain: Every client has its own stack, rules and history. Switching between them costs half a day.
Outcome: A project per client with its own agents, memory and rules. Nothing leaks from one client to another.
Yes. Arc OS is built around Anthropic Claude. We don't resell AI tokens. You bring your own. Claude Pro / Max / Teams: log in with OAuth. Anthropic API key (BYOK): paste your key, we pass requests through with no markup. Arc OS charges for the orchestration layer. Claude is yours.
Cursor and Claude Code are single AI assistants. Arc OS is a team: multiple role-specialized workers with shared project memory that persists between sessions. Notes (knowledge collections from PDFs, video, YouTube, and web) attach to issues. When a worker picks up a task, it reads the attached notes.
Chat history and API keys are encrypted at rest on our servers (AES-256-GCM, server-held key). To run your workers, our servers process messages in plaintext and send them to Anthropic (Claude). 2FA on all plans; GDPR Art.17 delete and Art.20 export. Servers in the EU.
Today: one owner per project. Project Collaboration (multi-user, invite by email, roles, shared worker memory) is in active development.
Not yet. Cloud only today, with data stored in the EU. A self-hosted edition is on the roadmap, earliest 2027.
Cowork is a personal assistant in Anthropic's cloud. Arc OS is an office for an AI team on a real server: named role-specialized workers, an issue tracker as the source of truth, an audit trail for every action, and hands in your infrastructure (repos, Docker, deploys). We don't compete with Claude. We build the home it works in.
You bring your own Claude subscription or API key with no markup, and Arc OS adds $0 (Free) or $4.99 (Starter) per month. We don't resell tokens, so we never profit from you burning more of them.
Bring your Claude subscription or Anthropic API key · 8 locales · CLI + web + Telegram · EU data.
arc-os.co: sign in requires JavaScript.