Frequently Asked Questions
The questions we hear most — from sign-up basics to architecture internals.
Product
Why not just use Claude's own app (Cowork)?
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 (developer, security, designer…) with their own skills and rules — not one generalist chat
- An issue tracker as the source of truth: a worker cannot start without a task, every action lands in an audit trail, commits without an issue reference are rejected
- Hands in your infrastructure: workers operate real repos, Docker, deploy pipelines on a server — not a sandbox
We don't compete with Claude — we build the home it works in. Every model upgrade makes Arc OS better for free.
How is Arc OS different from running several Claude Code sessions?
A session forgets everything when the terminal closes. An Arc OS worker is a persistent identity: role, skills, task history, chats. Workers coordinate through shared issues, wiki and semantic search — and every action is logged, so you always know who did what and why.
What does Arc OS cost on top of Claude?
You bring your own Claude subscription (Pro / Max / Teams via OAuth) or an Anthropic API key — passed through with no markup. Arc OS adds:
| Plan | Price | Includes |
|---|---|---|
| Free | $0 | 1 project, 5 workers |
| Starter | $4.99/mo | 5 projects, 25 workers |
| Cloud | $14.99/mo | 20 projects, 150 workers, dedicated cloud workspace |
We don't resell tokens, so we never profit from you burning more of them.
Is plain Claude cheaper then?
On its own — yes, by the price of an Arc OS plan. But the honest comparison is against the stack Arc OS replaces: an AI chat + an issue tracker + a knowledge base tool + your hours spent copy-pasting context between them.
Architecture
How are skills loaded? Doesn't that burn the context window?
Two levels. At start a worker receives only the intros of its role's skills. Then a context router picks the top-5 relevant skills per message; the full text of any skill is fetched on demand. Catalog: 140+ skills — table of contents always present, bodies lazy-loaded.
What is the CLAUDE.md hierarchy?
Three layers:
- Project CLAUDE.md in the repository — architecture, conventions, written by the team
- Arc OS block, injected automatically — system protocol, the worker's role and skills
- Dynamic layer — the active issue, issue-matched skills, and learnings: accumulated corrections that every future session inherits
How much token overhead does the system add?
The injected prefix (protocol, project state, active issue, skill intros) is roughly 10–15k tokens per session. With Anthropic prompt caching the stable prefix costs ~10% on repeat turns, so the real overhead is single-digit percent — and workers need fewer clarification round-trips, which usually nets out cheaper per task.
What's under the hood?
A deliberately simple stack: Bun + SQLite (WAL) + React + nginx — one server for the control plane, user cloud containers on a separate machine. Semantic search is self-hosted inside the same SQLite database (sqlite-vec + Cohere embeddings); no separate vector DB. New instances are provisioned with cloud-init in minutes.
Security & data
Can you see my code or chats?
Chat is end-to-end encrypted (AES-256-GCM; the master key is derived in your browser and never leaves it). Code lives in your isolated Docker container — we don't auto-read it. 2FA on all plans; GDPR Art. 17 cascade delete and Art. 20 export are built in.
Where is my data?
EU — on Hetzner (Finland/Germany), behind Cloudflare with authenticated origin pulls.