Arc OS — The Orchestration System for AI Teams

Your AI team for every project. One memory across CLI, web, and Telegram.

What Is Arc OS

Arc OS gives every project its own federated team of AI workers — role-specialized agents with dedicated skills, persistent memory, quality control, and self-improvement cycles. Built on Anthropic Claude: bring your Claude subscription or Anthropic API key, and Arc OS provides the orchestration layer on top.

One sentence: An orchestration system that deploys accountable AI worker teams — each with its own expertise, memory, and quality standards — reachable through a web CRM, the ARC CLI, and Telegram, with one shared project memory underneath.

How It Works

Three input channels feed the same orchestrator. Use whichever fits the moment — project memory stays identical across all three.

 You (CEO)
   │
   ├── CRM Dashboard (arc-os.co) ····· primary — all features
   ├── ARC CLI (arc <project>) ······· local dev, full context loaded
   └── Telegram ······················ mobile companion, per-worker topics
   │
   ▼
 Master Bot (orchestrator)
   │
   ├── Project A ──► own workers · skills · memory · evals
   ├── Project B ──► own workers · skills · memory · evals
   └── Project C ──► own workers · skills · memory · evals

Why three channels?

Channel	When you reach for it
CRM Dashboard	Most of the time — Workspace chat, Issues, Wiki, Notes, Skills, Timeline, Analytics, Billing
ARC CLI	Inside your IDE — `arc <project>` opens a worker session with full context; `arc issue / skill / kb / wiki / roadmap` without leaving the terminal
Telegram	On the road — one supergroup per project with a topic per worker (unified channel mode), notifications when a worker finishes

Each project is completely isolated:

CLAUDE.md — project constitution (tech stack, conventions, constraints)
Workers — role-specialized agents (13 preset roles + custom via Worker Studio)
Skills — loaded expertise from a 140+ skill library, plus project-local forks
learnings.md — accumulated corrections that survive restarts
Eval rules — automatic output validation per skill
Notes — knowledge collections (PDF, video, YouTube, web, audio, DOCX, images) indexed for RAG chat

Core Capabilities

1. CRM Dashboard (70+ API endpoints)

Workspace — chat and terminal workers with real-time streaming
Issues — tracker with kanban, phases, priorities, activity log
Notes — NotebookLM-style knowledge collections; sources attach to issues, workers read them when picking up tasks
Wiki + Knowledge Graph — per-project knowledge base with entity visualization
Skills — 140+ skill registry with Sage AI analyzer, benchmarks, and marketplace discovery
CEO Worker — cross-project assistant: sees all your projects, creates issues, searches knowledge (web /ceo page + arc ceo CLI)
Semantic search — self-hosted RAG (Cohere multilingual embeddings + sqlite-vec) across wiki, issues, skills, and notes
Timeline — DAW-style replay of all worker activity
Roadmap, Reports, Analytics, Billing, Project Settings

Auth: email/password or OAuth (Google, GitHub), optional 2FA/TOTP. 8 interface languages. Mobile responsive.

2. Federated Multi-Project Management

One Master Bot, unlimited projects. Each project runs isolated — no context bleed between your Odoo site and your SaaS app. Watchdog auto-restarts crashed workers.

3. Self-Improving Intelligence

Mechanism	What It Does
Binary Evals	Validates every response against declarative rules (6 rule types)
Context Router	Scores skills per message, injects only the top-5 relevant ones
Reflect Loop	Fix It / thumbs-down → persistent rule → injected into every future prompt
Karpathy Loop	Nightly metrics analysis → finds weak skills → proposes improvements → you approve with one tap

4. Worker System

13 preset roles (Consultant, Developer, Designer, Sentinel, Archivist, Product Owner, Analyst, Growth, CFO, Pitch Coach, Legal, Researcher, QA Engineer) plus fully custom workers via a 5-step wizard:

Property	Options
Model	Sonnet / Opus / Haiku
Tools	Per-worker tool scope (read-only → full repo access)
Skills	Pinned context assets from the skill library
Avatar & color	Role presets or custom upload
System prompt	Preset, AI-generated, or custom

5. Cloud or Local

Standard Cloud — your own Docker container on EU infrastructure (Hetzner); GitHub App auto-creates private repos per project
Local Bridge — zero-dependency binary: workers run on your machine using your local Claude session
Cloud chat routing — on the cloud plan, chat messages execute inside your container with your Claude login; no key required

6. Telegram Command Center

Unified channel mode: one bot per project, one supergroup with a topic per worker. Messages route by topic; "worker finished" notifications land in the right thread. Inline controls on every response: [STOP] [PAUSE] [Fix It] [👍] [👎].

7. ARC CLI

arc <project>             # boot a worker session with full context
arc <project> <worker>    # boot a specific worker
arc ceo                   # cross-project CEO assistant REPL
arc issues                # list project issues
arc issue create / log / take / switch
arc skill <name>          # load skill instructions
arc kb search "..."       # semantic search across project knowledge
arc wiki update / roadmap sync / report / sessions / retro

8. Security & Compliance

Zero-knowledge E2EE chat (AES-256-GCM at rest, PBKDF2 master key, recovery keys)
2FA/TOTP on all plans; HIBP password check at registration
GDPR: Art. 17 cascade delete, Art. 20 data export
AES-256-GCM secrets vault; mTLS origin pulls; isolated per-user containers

Numbers

Metric	Value
API endpoints	70+ across 19 domain modules
Skill library	140+ global skills
Worker role presets	13
Interface languages	8 (EN, UK, DE, ES, FR, PL, RU, pt-BR)
Note source types	8 (PDF, video, YouTube, web, audio, DOCX, TXT, images)
Eval rule types	6
Encryption	Zero-knowledge E2EE (PBKDF2 + AES-256-GCM, recovery keys)
Deploy model	SaaS at arc-os.co — early access

Who Is This For

Solo developers who lose hours re-explaining the same project to different AI tools
Technical founders running multiple products who need isolated AI teams per project
Agencies managing client projects with different stacks — per-client memory and skills
Teams that want AI with memory — corrections stick, mistakes don't repeat

Get Started

Early access: registration currently requires an invite code (arc-XXXX-XXXX) or a waitlist spot — request access at arc-os.co.

Via CRM Dashboard (recommended)

Open https://arc-os.co
Sign in with Google/GitHub or email → paste your invite code
The onboarding wizard creates your first project — pick a blueprint, connect your Claude subscription or drop an Anthropic API key, ~60 seconds to first message

Via ARC CLI (developers)

arc login --server https://arc-os.co
arc <project-name>        # boots a worker session with project context

Via Telegram (companion)

Connect a bot under Project Settings → Channels — one supergroup per project, a topic per worker.

Full setup: User Guide · ARC CLI Reference · Creating Skills