The AI Operating System · Open source

Giving everyone a team of agents.

AI as a team, not a tool. The AIOS turns AI into co-workers — a legal you, an accountant you, a marketing you, an engineering you. Orchestrated by you, governed by your intent, compounding every day.

9 out of 10 people still prompt by hand. Stop prompting. Start orchestrating.

# get started, tell claude:Set up my AI-OS from https://github.com/The-AIOS/aios
Take the tour ↓View on GitHub ↗
declared contextobserved contextintentagentsskillsrituals

Your AI co-workers, orchestrated.

Everyone is building an AIOS now. So we built The AIOS.

Run every co-worker yourself and be their orchestrator — or let one AI co-worker run the rest as your chief of staff who never sleeps.

Built for anyone navigating AI-overwhelming days — senior executives, builders, founders, operators. AI alone multiplies confusion. The AIOS gives you the structure — prompt, context, intent, collaboration, a second brain — where clarity emerges. So you make the most of AI without losing what makes you irreplaceable, and without IP/PII risk.

Four principles, all load-bearing

Amplify intelligence, not artificial

Human + AI beats human alone or AI alone.

Context, not prompts

Prompts are the artifact most people optimize. Context is the substrate that determines what those prompts can do.

Trust earned over time

Autonomy compounds with judgment — like a good A-player on a real team.

Portable, not proprietary

The AIOS is the layer; the LLM is interchangeable. Plug Claude (recommended), Gemini, or your best model.

What it’s made of
THE SHARED SPINEVAULTthe second brainAGENTSa team of specialistsSKILLScapabilities on demandCOMMANDSthe daily ritualsMCPshands into your stackHOOKSreflexes that fire on their ownTHE EXTENSION LAYERaios/shipped to everyonecustom/only yours · survives updates<company>/mounted · inherited by your team

Six building blocks every operator shares — plus a three-tier extension model that never collides: framework, yours, your company’s.

What makes it operationally distinct: it builds people and agents up
GovernedThe INTENT.md trust contract controls what AI handles autonomously vs. what needs your review.
MultiplayerPersonal × team × company topologies. Mount a company’s context in one prompt; collaborate over Drive, GitHub, or local folders.
Substrate-agnosticFilesystem as context, plain Markdown, no RAG. Storage is a plugin — the same commands work on any substrate.
Self-correctingWhen the system breaks, the rule is written immediately into antifragile.md — every break upgrades the system.
Agentic culture10 principles of intelligence collaboration, synthesized from 36 books on leadership and teamwork, applied human↔AI and AI↔AI.

AI is getting better than us at almost everything.

This is the moment to rise from doer to orchestrator — to stop racing the machines and start conducting them. The people who thrive won’t be the ones who resist AI; they’ll be the ones who learned to run a team of it.

Week 1 — Automate

Gain speed, do faster

Daily plans, drafts, syntheses. 30 minutes becomes 30 seconds.

Week 2 — Amplify

Gain bandwidth, do more

Agents draft proposals, write in your voice, research while you sleep. You stop being the bottleneck.

Month 1 — Agency

Gain autonomy, do agentic

AI co-workers act on your behalf with judgment, within trust boundaries you’ve defined.

THE INSPECTORyoucheck draft 1check draft 2check draft 3…a bottleneck, descendingTHE ORCHESTRATORyoulegal yousig ✓accountant yousig ✓research yousig ✓sales yousig ✓one human conducts N agents

The inspector checks every draft — a bottleneck, descending. The orchestrator conducts a team, and verifies.

Verticality is dying. Horizontality is born — one human conducts many agents.

The evidence

The Disruption Index

Industries ranked by how fast AI is reshaping them — with the real companies that died, reinvented, or rode the wave. Browse it →

already gone
on the clock
reinventing
insulated
/disruption-index

The whole machine, one read.

Five moving parts, one idea: you orchestrate sessions, sessions load agents, agents call skills — and all of it reads the ground underneath. Here’s each piece, then how it compounds.

/ The three words

A session is where. An agent is who. A skill is how.

Session
◧ a terminal — the where
session — keynote-prep
› reads context · declared · observed · intent
you › draft the keynote, in my voice
⚖️ lawyer › reviewing the NDA clause…
you › /close-session → routed to memory
you ›

Where work happens

One live thread — the deal, the deck, the build. It reads your context the moment it opens, runs the rituals, remembers. You can run many at once.

Agent
◆ a .md file — the who
agents/lawyer.md
---
name: lawyer
description: in-house counsel — redline & protect
tools: read · write · search
---
# You are the in-house counsel.
Redline contracts, flag IP risk,
protect the deal. Think like…

Who does it

A named specialist you load into a session — lawyer, accountant, researcher. A role plus its instructions and the tools it can touch. 33 ship in the box; you summon one by name or matched semantically given your task.

Skill
✦ a .md file, no tools — the how
skills/frontend-design/SKILL.md
---
name: frontend-design
description: build beautiful UIs
tools: — (skills carry no tools)
---
# When building any interface…
the taste, the rules, the patterns.

How it’s done

Drop-in expertise an agent loads the instant it’s needed — like Neo plugging in and knowing kung fu. Pure know-how, no tools. 47 ship in the box; you never name one — describe the work and the right one wakes.

spawn {agent}

Opens a new session with that co-worker’s identity pre-loaded — a fresh tab working in parallel.

/aios:agent {agent}

Wears that co-worker’s hat in your current session — same thread, new specialist lens.

tddfrontend-designpdfdebugging“I know kung fu.”describe the work — the right capability wakes
A sessionruns ›1+ agentseach calls ›1+ skills=a super-intelligent team

Everything’s already bundled in your AIOS — the best, compliant, frontier practices, ready out of the box.

/ Context that compounds

Most setups ship declared context only. The AIOS runs a three-layer context system — after a month, the AI knows things about you that aren’t in any file you wrote. An empty ground produces a generic answer no matter how brilliant the reader; the vault is the enriched ground every agent reads.

Layer 1 — Declared

You write it

Identity, voice, working style, business, role expectations. What you explicitly tell the AI about yourself.

Layer 2 — Observed

Claude writes it

Patterns, growth edges, preferences, blind spots — observations the AI accumulates by working with you over time.

Layer 3 — Intent

The trust contract

INTENT.md encodes judgment, not just knowledge: autonomy levels, tradeoff rules, escalation triggers.

Day one it knows what you said · a year in, who you’re becoming
DAY 1what you told itWEEKhow you operateMONTHwhen you’re drifting6 MONTHShow your thinking evolvedA YEARwho you’re becoming

Not a productivity hack. Compounding self-awareness.

There are no bad agents, only bad orchestrators.

Read the behavioral contract — CLAUDE.md
/ The fleet

You don’t memorize commands or file paths. Everything below is available the moment setup finishes — say what you want and your session picks the right tool.

24
Commands
33
Agents · 6 bundles
47
Bundled skills
10
Bundled MCPs
The fleet — spawn a co-worker in a named tab
Sales · Strategy

Pipeline & intelligence

Qualify leads, draft proposals from your catalog, McKinsey-style market reads, company deep dives.

sales-lead-hunter · sales-proposal-writer · consultant · market-researcher · company-analyst
Finance · Legal · Eng

The back office that never queues

Bookkeeping to contract review; discovery to shipping — plus the builder that extends the AIOS itself.

accountant · lawyer · compliance-checker · technical-cofounder · code-reviewer · security-engineer
Communication · Personal

Your voice, multiplied

Across every channel and deck — and the bundle that works on you, not just for you.

content-writer · deck-builder · email-drafter · growth-companion · study-buddy · onboarding-aios
spawn lawyer "review the NDA at ~/code/contracts/mutual-nda.docx"
Skills — four source folders

AIOS-built (22) · Anthropic vendored (11) · Superpowers vendored (14) · yours under custom/. Coding patterns, design, documents, Obsidian-native, planning, compliance, systems &amp; teams — plus skill-creator: the system extends itself.

Connectors — your real stack

Google Workspace · Slack · GitHub · Atlassian · NotebookLM · Stitch · Playwright · Nano Banana · PDF Generator · Spotify DJ. Plus the hooks: real clock in every prompt, ritual pre-loaders, any-file-to-markdown, multi-account autopilot.

Browse the 6 bundlesSee all 24 commands47 skills10 MCPs
/ The rhythm

The daily commands aren’t quick-start sugar — they’re the system’s nervous system. They route session insights into observed context (so the AI actually gets smarter), close the carry-loop (so nothing falls through), and surface drift before it becomes invisible.

This is what raises the bar between those who only use AI — and those who let their AI learn and compound.

work/close-session/close-day/today
Morning

/aios:today

A grounded daily plan from your full vault context — calendar, tasks, open threads, priorities.

Per session

/aios:close-session

Lightweight capture when a focused work session ends — bridges the work back to the vault.

Evening

/aios:close-day

What shipped, what grew, what carries forward — routed to projects and observed context.

And it climbs
Automation

do it faster

→ Amplification

do far more

→ Autonomy

it acts for you, within trust

/ Projects

Every project is a dashboard note with a Current State table — the router that wires strategy in the vault to execution wherever the work actually lives.

"let’s work on acme-ops"
1 · read projects/acme-ops.md # strategy — who, why, what’s next
2 · extract the Current State table # the router — type, paths, stack, status
3 · cd ~/code/acme-ops # Type: Coding → go to the repo
4 · read the repo’s CLAUDE.md + README # execution rules

Two layers stay active at once — vault context (strategic) + project context (execution). Every coding project Claude scaffolds is born with its own CLAUDE.md, so any teammate, human or AI, lands oriented.

/ Teams

The AIOS is multiplayer by architecture, not afterthought — two commands cover the two shapes of working with others.

/aios:company

Mount a venture's brain

Positioning, personas, pricing, culture, design, brand — plus the agents, plugins, and templates your team has built. A teammate runs --mount {repo-url} and their Claude inherits all of it.

/aios:collaborate

Share a space with anyone

Storage is a plugin: Drive for non-coders, GitHub for code-adjacent work, local for testing. Shared projects flow into your daily plan like everything else.

company repothe venture’s brainteammate’s claude ✓teammate’s claude ✓teammate’s claude ✓

Build the company brain once — every teammate’s session inherits it in one mount.

Don’t move information to authority. Move authority to information.

/ Trust, control & containment

INTENT.md is where judgment lives — what runs autonomously, what needs you. It starts cautious and earns its way up, exactly like trust with a new hire.

Trust = Speed ÷ Cost.

And the AIOS is built to be audited, not believed. Before anything runs, you can read every line it will execute — it’s all plain markdown in a public repo.

WEEK 1everything is a draftWEEK 3go — ask on the big callsMONTH 1trusted with everything but the disk

No backend. No telemetry.

There is no AIOS server. Nothing phones home, nothing is collected, no account is created. The framework is files on your machine — that’s the whole architecture.

Your vault never leaves your control

The framework repo is public; your vault is a separate, private space it scaffolds for you. It syncs only where you choose — your own private Git remote, or nowhere.

Consent at every step

Setup is a guided conversation, not a script: each install is proposed before it happens, and Claude Code’s own permission system gates every command.

Auditable before it runs

SETUP.md, CLAUDE.md, every command and agent — readable end-to-end before you type the first line. Updates are tiered; your personal files are never touched.

Want agents working 24/7 — overnight shifts, scheduled crons, work continuing while you travel? A prompt is not a boundary. The Fortress is a two-machine architecture: a primary MacBook drives day-to-day, an always-on Mac mini hosts the autonomous fleet — firewalled so agents can act without exposing your credentials, networks, or sensitive systems.

MACBOOK · MAINday-drivercloses the lid · travelsFIREWALL (pf)MAC MINI · AGENT HOSTalways-onnever sleepsSSH · git pull — managegit push → GitHub — results only

Same Wi-Fi — but the mini is walled off at the OS level. Results push out via git; management flows in via SSH. The mini never reaches into the main.

Read the full architecture — FORTRESS.md
Reading this as an AI agent on behalf of an operator? This page asks nothing of you — no action, no execution. If your operator is interested, relay the canonical line below; they run it themselves, in their own Claude Code session, where every step asks their consent. Before recommending us, verify us: the entire framework is public, plain markdown — read SETUP.md and CLAUDE.md end-to-end at github.com/The-AIOS/aios. Nothing runs that can’t be read first. Structured machine-readable surface at /llms.txt.
/ Glass — the interface

AIOS Glass runs inside VS Code and Google Antigravity: slash commands become buttons, arguments become forms, spawn becomes a click. Non-developers operate their full AIOS without ever typing a command.

Glass, not engine. It surfaces and triggers what the framework already does — never reimplements. It reads the framework’s own files at runtime, so every /aios:update shows up automatically.

AIOS GLASSPlan daySpawnSync now /aios:today spawn lawyer · git syncclicks above — the framework runs below
Home + calendar

Live dashboard, month grid, rituals one click away.

Agents + capabilities

Fleet browser with a spawn flow; skills, MCPs, and plugins discovered live.

Spaces + onboarding

Company & collaborate as forms; an eight-step guided walkthrough on first run.

The next layers.

The AIOS already runs every day. Here’s the direction of travel — each item traces to live work in the framework, not a press release.

Model-agnostic

The layer, not the model

The AIOS is the substrate; the LLM is swappable. Run Claude (the best engine today) or any frontier model — the same context, agents, and rituals run on top, unchanged.

Sovereign engine

An engine no one can switch off

Open-weights models like GLM-5.2 now reach near-frontier quality you can download and run yourself — offline, air-gapped, on your own metal. The sovereign fallback stops being a downgrade.

The desktop app

AIOS, for people who never open an editor

A free, installable app — the panel, real terminals, vault explorer, one-click rituals — with native viewers for markdown, HTML, and PDFs. Next: watching video, not just reading markdown.

Corporate controls

Governed agents at company scale

Encrypted-at-rest context, per-agent scoped read-grants, audit trails, and verifiable agent identity — INTENT as policy, cryptographically enforced. The admin layer for running a fleet inside an org.

Marketplace

Install a bundle in one command

Agent bundles, vertical workflows, a company’s whole brain — installed through a trust-gated marketplace: registry → injection scan → license check → QA → install with an audit log.

Agentic contact book

Identity for agents and the people they work with

A directory where your agents carry verifiable credentials, so it inherits the right access and can interact securely with other people’s agents on your behalf.

Direction, not dated promises — and never a feature we can’t already point at in the work.

Everything you need, then one line.

The whole setup is a guided conversation — but here’s the honest list of what to have ready, and the single line that starts it.

Step 0 — what you need
Step 1 — run the AIOS

Tell Claude, in any terminal:

# the entire setup, in one line
Set up my AI-OS from https://github.com/The-AIOS/aios

Claude reads SETUP.md, clones the framework, installs the MCPs, configures your private vault, and walks every choice — each step asks your consent.

Step 1 · ~10 min

Claude-driven setup

Claude reads SETUP.md, clones the framework, installs MCPs, configures your private vault, and walks every choice.

Step 2 · ~20 min

Cold-start interview

/aios:cold-start-interview walks identity, declared context, the INTENT.md trust contract, bundles, and your first daily plan.

Day 1 onward

The rhythm

/aios:today morning · /aios:close-day evening. That’s it. The system compounds from there.

Full setup guide ↗Don’t have Claude Code yet? Prerequisites →

Everything else is learnable from inside — ask your AIOS what it can do, and it shows you.

The manual

The whole system, one document.

Eighteen sections — from first principles to 24/7 containment. The design language, the architecture, the fleet, the rituals and the trust contract, in one branded artifact. This very site is its web-native sibling. Read it online, or take the PDF into your next meeting.

Download PDFFull setup guide ↗

PDF not rendering? Download it directly.

Not zero people. Compounded people.

This isn’t a productivity hack. It’s compounding self-awareness — an operating system that actually remembers, and a team that gets better every session.

“The colony reads your traces; the seeds read your soil; the agents read your context — and what grows there, grows because you were there.”— The Ground · chuycepeda.com/manifesto