MIT · Windows first · Linux, macOS, WSL, Docker

Open-source AI agents, installed, wired and run as one system.

One CLI installs opencode, Hermes, OpenClaw, CrewAI, Claude Code, Codex and Paperclip, points them all at one model config, and connects them through a hub on :8787. From there you talk to any of them, hand a job to the whole team, or file it as a ticket and let the org chart handle it.

PS> irm https://raw.githubusercontent.com/ZDStudios/AIOS/main/install.ps1 | iex

$ curl -fsSL https://raw.githubusercontent.com/ZDStudios/AIOS/main/install.sh | bash

$ docker pull ghcr.io/zdstudios/aios:latest
# or build it with compose:
$ git clone https://github.com/ZDStudios/AIOS.git && cd AIOS
$ cp .env.example .env && docker compose up -d

The installer asks for admin once, installs uv, bun, pnpm and Node into your profile, clones to ~/AIOS and runs aios setup. No API key? Pick your Claude or ChatGPT subscription in the setup wizard instead.

The AI OS hub: chat view with conversation list, message composer with target chips (Brain, Team, Swarm, Company, Codex and others), and a live list of running agents. — **The hub's chat view.** The chips under the transcript choose who answers: a single agent, the Brain, the whole team, a swarm, or the company. The right rail shows which services are up.

What's inside

Each project is the real upstream, not a fork. AIOS installs it, renders its config from one .env and one aios.config.yaml, and keeps it updated under review.

Project	What it does here	Runs as	Upstream
Agents and services
opencode	Coding agent. The hub sends it code tasks over its HTTP API.	server :4096	anomalyco/opencode
Hermes	Autonomous agent with memory, cron and a learning loop. Its dashboard is embedded in the hub.	server :9119	NousResearch/hermes-agent
OpenClaw	Messaging gateway for 28 channels (Telegram, Discord, Slack, WhatsApp…) and the plugin host.	gateway :18789	openclaw/openclaw
CrewAI	Role-based crews, with an `ask_peer` tool that calls the other agents.	server :4788	crewaiinc/crewai
Claude Code	The Claude CLI exposed as an OpenAI-compatible API, so a Pro/Max subscription can power every agent.	API :8000	claude-code-api
Codex	The Codex CLI as an OpenAI-compatible API, on your ChatGPT login.	API :8010	openai/codex
Paperclip	The company layer: org chart, tickets, heartbeats, budgets. AIOS hires its agents into it.	server :3100	paperclipai/paperclip
Mounted into every agent
Fabric	255 task patterns (`summarize`, `extract_wisdom`, `analyze_claims`…) run on your own model.	prompts	danielmiessler/fabric
Caveman	Terse-output mode. Roughly 65% fewer output tokens; code and errors stay exact.	prompt overlay	JuliusBrussee/caveman
Ponytail	Build-less mode. Pushes agents to the smallest change that works.	prompt overlay	DietrichGebert/ponytail
Ruflo	Meta-harness with 314+ MCP tools, driven through `aios ruflo` for heavy jobs.	npx, on demand	ruvnet/ruflo
LifeOS	Shared identity and skills.	skills	danielmiessler/LifeOS
OpenUI	Generative-UI context. Agents can answer with live widgets; openclaw-os renders them.	context + plugin	openui.com

How work moves

Three ways to hand something off, depending on how big it is and whether you want a record of it.

Ask

Pick a target and send. Small questions go straight to one agent. Auto routes by content: code to opencode, research to CrewAI.

Swarm

A big request becomes a graph of subtasks. Independent ones run in parallel, each on the agent that suits it. Mid-task, an agent can write ASK opencode: … and gets a real answer back.

Company

File a ticket. The CEO delegates down the org chart, the assignee does the work, and the result lands on the ticket with a status you can audit later.

A ticket, end to end

Paperclip owns the state; the hub does the work inside the heartbeat request, so Paperclip's run log matches what actually happened.

You file THE-12From the Company view, the 🏢 Company chat chip, or aios paperclip task "…".
Paperclip wakes the assigneeA heartbeat hits POST /api/paperclip/heartbeat on the hub.
The hub checks it out and routes itTo Brain, opencode or CrewAI. Claude and Codex run on Paperclip's own local adapters.
The result is postedAs a comment. The ticket goes to done, or to blocked with the reason and the board as owner.

AIOS Brain CEO http → hub ├ opencode CTO http → hub ├ CrewAI Head of Research http → hub ├ Claude Senior Engineer claude_local └ Codex Engineer codex_local

Company view in the hub: counts of queued, working, blocked and done tickets, the org chart, a new-ticket form and a four-column ticket board. — **Company view.** Org chart, a board that refreshes while tickets are in progress, and each ticket's comment thread. Paperclip's own UI is one tab over.

What else the hub does

The parts you notice after the first day.

Live actions: Every shell command and tool call an agent makes streams into the chat as it happens. One toggle in the header turns the feed on or off.
Supervised updates: AIOS watches each project's upstream. When one moves, an agent reads the commits and changed files and answers SAFE, RISKY or BLOCK. Prompt-only projects apply on SAFE; anything with a dependency tree waits for you. aios updates --rollback restores the previous copy.
Self-healing: A watchdog health-checks every service. A dead one is restarted; if that fails, a healthy agent reads the logs and writes up the cause.
Reads the web: YouTube transcripts and chapters, Reddit threads with comment trees, Hacker News, any page as clean text, and DuckDuckGo search. No keys. Paste a link or use /yt, /reddit, /hn, /web.
Subscriptions, not keys: Run everything on a Claude Pro/Max or ChatGPT plan through the local bridges. The hub is itself an OpenAI-compatible API at /v1, where the "models" are the agents.
Agents restyle the UI: Ask for a different accent or a new panel and the agent calls /api/dashboard, never the HTML. Input is validated (no @import, no remote url()) and one Reset undoes all of it.
Themes and plugins: Install a theme by pasting plain CSS. The plugin store lists channels, Ruflo plugins and skills, installable from the dashboard.
Memory and schedules: A memory sub-agent runs on every turn. Task Brain handles cron and interval jobs; TaskFlow chains agents and resumes from the last committed step after a crash.
Skills that accumulate: After a task that ran commands, a curator decides whether it taught something reusable and writes a SKILL.md that every agent then mounts.

Status view: a grid of service cards (hub, opencode, hermes, openclaw, crewai, claudecode, codex, paperclip, openclaw-os) each marked UP with its local URL, above the self-healing log. — **Status.** Every service with its URL and restart/stop controls, above the watchdog's incident log.

Security model

The agents have a shell on your machine; that is the point of the project. So the control plane is locked down and the destructive paths need a password the agents can't read.

The hub

Loopback is trusted, nothing else is. Every non-loopback request needs AIOS_HUB_TOKEN, minted at setup.
CSRF closed. Cross-origin requests are rejected without the token; CORS never echoes *.
DNS rebinding closed. Host must be an IP literal or a known local name.
Guardrails. The shared shell refuses rm -rf /, mkfs, dd to a block device, fork bombs and shutdown.
Audit log. Every command any agent runs is recorded in .aios/aios.db and shown in Control & Audit.

The master password

Never stored. Only scrypt (n=2¹⁵, r=8) of it is. There is nothing to read back.
Fails closed. No vault, or one that doesn't verify, refuses the operation. Deleting the file locks you out; it doesn't let an agent in.
Tamper-evident. Mirrored outside the repo and checksummed; a mismatch blocks protected operations.
Off-limits to the agents' shell. The guarded shell won't touch the vault paths.
Gates aios uninstall, disabling guardrails and enabling full control.

Stated plainly: an agent that writes files by some route other than the guarded shell could still delete the vault. It can't recover the password from it, and deleting it grants nothing. If you'd rather contain all of this, run the Docker image; full control then means full control of a container you can throw away.

Commands

Everything the dashboard does is also a command. aios help lists the rest.

Command	What it does
aios setup	Install toolchains and dependencies, run the model wizard, render configs, mount skills. Safe to re-run.
aios start \| stop \| restart [svc]	Bring services up in order and wait on health checks. Kills the whole process tree on stop.
aios status · aios doctor	Service table; diagnosis with the exact fix for each problem.
aios paperclip setup \| task \| open	Hire the agents into the company, file a ticket, open Paperclip's board.
aios updates [--apply P \| --rollback P]	Supervised upstream updates with an agent's verdict on each.
aios web youtube \| reddit \| hn \| fetch	The web readers, from the terminal.
aios ruflo …	Pass-through to `npx ruflo@latest`; `plugins` and `mcp-register` helpers.
aios exec <cmd>	Run a command through the agents' guardrailed, audited shell.
aios channels · aios migrate	List the 28 channels; import an existing OpenClaw install.
aios password set · aios uninstall	Set the master password; remove AIOS (asks for it).
aios url · aios token	Print the hub URL (and WSL address); print or rotate the hub token.