AutoYou Server 发布：在自有电脑上自托管 AI 模型与智能体

You are not the product. You are the owner.
A self-hosted AI server for the computer you already own.

Quick start · How it works · Capabilities · Apps · Architecture · Docs · Contributing · License · Security

AutoYou Server runs on your own Windows, macOS, or Linux machine. It hosts your AI models and agents, speech, voice and video, notes and memory, and a local web console, and it lets your phone, other computers, and browsers reach it over encrypted peer-to-peer connections. Conversations, files, and agent activity are stored on that machine, not in a hosted AutoYou service.

AutoYou client apps are distributed separately. Server-backed features need a server that you run or compile.

Quick start

You need Git and Python 3.10 or newer. Dependencies can require a newer Python or a particular platform; see the dependency profiles.

1. Get the server

git clone https://github.com/autoyou-ai/AutoYou-Server.git
cd AutoYou-Server

2. Run the launcher

Windows:

.\run_autoyou.bat --profile base

macOS or Linux:

./run_autoyou.sh --profile base

The launcher bootstraps a Python environment, installs dependencies from package indexes, and may download models and native tools. Read the scripts first and run them in an environment you can safely modify.

3. Open the console

Go to http://127.0.0.1:8001/, choose a unique server password, and follow the setup to pick your AI, enable the tools you want, and pair a device.

For unattended first setup, provide AUTOYOU_SERVER_PASSWORD through your environment or a secret manager. Keep real passwords out of shared commands, issues, and screenshots.

Choose a profile

Profiles select which dependency sets are installed. They are cumulative, and the launcher uses recommended when --profile is omitted.

The root launchers also request the internet (browser) component, including for base. Profiles choose dependencies only; integrations still need configuration, and optional features may download models or native tools and require provider accounts. See requirements/README.md for platform limits and the current dependency review.

How AutoYou works

AutoYou is built around one idea: the server is your own computer, and devices talk to it directly. The sections below explain what that means for connectivity, for the services involved, for inspectability, and for security, including where the guarantees stop.

Peer to peer, without port forwarding

Apps and the server connect with WebRTC. The two sides exchange a short pairing handshake, then open an encrypted link (DTLS, with SRTP for media) that carries chat, voice, video, screen sharing, and proxied web traffic. Most home and mobile networks can form that link without opening router ports. When a network is too strict, the connection falls back to a relay that forwards the same encrypted packets.

flowchart LR
    app["AutoYou app<br/>phone, desktop or browser"]
    server["AutoYou Server<br/>your computer"]
    rendezvous["Rendezvous (optional)<br/>Cloud Pair"]
    relay["Relay (optional)<br/>forwards encrypted packets"]

    app <-->|"encrypted WebRTC, direct when the network allows"| server
    app -.->|"signaling"| rendezvous
    rendezvous -.->|"signaling"| server
    app -.->|"fallback path"| relay
    relay -.->|"fallback path"| server

Loading

You choose how a device finds your computer:

Peer Link is separate from pairing: it connects one AutoYou app to another app, and the receiving app approves each link. See docs/technical/pairing-protocols.md.

Pairing messages (offer, answer, and ICE candidates) are encrypted under a key from a password-authenticated key exchange: a CPace-family construction over ristretto255 using libsodium. A relay or messaging bridge that carries them cannot read them, and a captured transcript cannot be used to test password guesses offline. The module documents where it differs from the registered ciphersuite: shared/pairing_cpace.py.

Serverless, stated precisely

"Serverless" here means that no AutoYou-operated application server is needed to run your assistant or to hold your data. You operate one server, your own. The server is designed to run on a local network without an AutoYou Cloud connection (details). Everything else in this table is either a default you can change or an optional helper, and each has its own trust boundary:

The connection-helper documentation states that AutoYou Cloud does not decrypt message content, voice audio, files, or browsing content. Peers and connection services can still learn network addresses and timing, because WebRTC does not hide your public IP from the other side.

Public STUN is on by default. Unless you configure otherwise, the server advertises Google's public STUN servers so WebRTC can discover its public address, and this applies to Local Pair too. To keep that lookup off the internet, set AUTOYOU_DISABLE_PUBLIC_STUN=1 before starting the server, or point it at your own STUN and TURN host with AUTOYOU_LOCAL_STUNTURN_HOST (plus AUTOYOU_LOCAL_TURN_USERNAME and AUTOYOU_LOCAL_TURN_PASSWORD for TURN). Without any STUN or TURN server, connections across NAT may fail.

Transparent by construction

• Readable source. The code that handles your data is in this repository. You can read it, build it, and run it. The license is source-available, not OSI-approved open source.
• Notices and dependencies. THIRD-PARTY-NOTICES.md and the dependency review record what the project includes. The review includes a reproducible advisory check: python scripts/audit_dependency_advisories.py --all-manifests.
• Automated checks. Pull requests run the public checks, CodeQL analysis, and dependency advisory scans in GitHub Actions.
• Visible state. The admin console's live view shows connected devices, running tasks, and enabled integrations, so you can see what the server is doing and who is connected.
• Stated limits. The security policy and the dependency review say what they do not cover. Published source and inventories support inspection; they do not certify an installation as secure.

Secure, with the boundaries spelled out

Encryption protects the connection. It does not protect everything around it:

• Endpoints. Malware on the server computer or on a paired device has the access that endpoint has.
• Agents. Agents run with the server's permissions. Browser automation, file access, desktop control, and command execution are powerful; enable only what you need and review what agents do.
• Third parties. Cloud model providers, messaging bridges, and public tunnels have their own data flows and terms. Review each before enabling it.
• Operations. Running a server makes you responsible for updates, backups, and credentials.

Report suspected vulnerabilities privately through SECURITY.md.

Capabilities

Features depend on the selected profile, hardware, permissions, and configured integrations.

Client apps

AutoYou Server runs on your computer. The apps reach it from other places, and the phone apps also carry a small local AI for when no computer is available.

Pairing modes	Chat with your own AI	Nearby Lobbies	Local AI on the phone

• Local AI without a computer. Local chat runs on the phone with LFM 2.5 350M, Gemini Nano, or Apple Intelligence, depending on the device.
• Lobbies. Host or join nearby rooms over Wi-Fi and Bluetooth, with chat and audio or video for up to six guests and the host.
• Peer Link. Connect one app to another. The receiving app approves each link, and optional AI replies answer using only what you told it.
• Remote desktop. View and control your computer's screen with touch and hardware modifier keys.

The ecosystem overview covers every client.

Admin console

The local console brings setup, models, agents, connections, and security controls together. These screenshots use example account and network values; statuses reflect the captured setup. Select a preview to open the full image.

Server overview	Devices and activity	Access and security

Architecture

• Pair. Auto Pair exchanges an encrypted pairing message through a Telegram, Signal, or WhatsApp workflow you set up, or pairing happens over your local network.
• Connect. Apps reach the server over an encrypted WebRTC link for chat, voice, and video. MCP clients connect through the MCP bridge. Agent Apps use the same link as HTTP over the SCTP data channel.
• Think. The intent router connects agent harnesses: Google ADK runs the native agents, and OpenClaw, Hermes Agent, and Odysseus are bridged. LiteLLM sends model calls to Ollama on your machine or to optional cloud providers.
• Listen, speak, remember. faster-whisper transcribes, system voices speak (EmotiVoice is optional), and memory lives in SQLite (Cognee is optional).

Dashed boxes in the diagram are optional add-ons or bridged projects. The diagram is generated by scripts/build_architecture_diagram.py.

Other ways to run

Docker

Provide AUTOYOU_SERVER_PASSWORD securely, then run:

docker compose up --build

The Compose file binds published ports to host loopback and keeps configuration in a named volume. Check how your Docker version publishes ports before relying on loopback isolation.

To disable build-time and runtime Tunnelmole downloads, set both AUTOYOU_SKIP_TUNNELMOLE_DOWNLOAD=1 and AUTOYOU_NO_DOWNLOAD_TUNNELMOLE=1 before building and starting. Features that need it then require a separately supplied binary.

Compile an unofficial local build

From the repository root:

# Windows
.\servers\windows\build-all.ps1 -Configuration Debug

# macOS
./servers/macos/build-all.sh --no-sign --dev

# Linux / WSL
./servers/wsl/build-backend.sh --unofficial

These routes do not require an AutoYou account or official release approval. They may install or reconcile dependencies. Review the license and notices when prompted. For unattended acknowledgement after review, pass -AcceptTerms on Windows or --accept-terms on macOS and Linux.

Prerequisites, output locations, and known limitations: Windows · macOS · Linux / WSL.

Documentation

Contributing

Start with the built-in agents, the Agent Builder, or the custom agent tutorial.

• Build an agent, extension, integration, or web interface.
• Improve setup, accessibility, platform support, documentation, or a reproducible bug.
• Share a demo, a guide, or a clearly labeled unofficial build that follows the license.

Read CONTRIBUTING.md and the Code of Conduct before submitting work. Independent original code can use your chosen license, including MIT or Apache-2.0. Changes to existing AutoYou code keep its applicable license.

License

AutoYou Server is source-available under the server license. It is not currently an OSI-approved open-source project, and the definitions in LICENSE control.

Preserve notices, identify unofficial builds, and meet the obligations of included components. See THIRD-PARTY-NOTICES.md and the trademark policy.

OpenStorey allocates at least 15% of qualifying net receipts to the contributor pool, awarded under the pool's review and eligibility rules. The funding roadmap includes a $5M server open-source milestone; a milestone does not itself change the license of a copy you have. Details: docs/legal/open-source-commitment.md.

Support and security

• Questions and setup help: SUPPORT.md and the community.
• Bugs and focused proposals: open an issue.
• Vulnerabilities: report privately as described in SECURITY.md. Do not use public issues for undisclosed vulnerabilities.

Agents run with the server's permissions, so review their actions, enable only what you need, and keep backups.

Copyright © 2026 OpenStorey LLC and contributors.