Pagis 开源多智能体助手：每个 Agent 拥有独立记忆与容器

A staff of AI Agents that work for you, each with a job, a memory and a computer of its own.

Pagis gives you a staff of Agents: AI helpers that work as your virtual assistants. Each Agent has a job, a personality, its own memory and its own computer, a container with a browser and a terminal. You work with them in a chat app, in direct messages, group channels and threads. The interface calls an Agent a sprite. One Rust daemon serves one person on a laptop or a team on a server. It sends model requests to the providers that you configure (Anthropic, OpenAI or OpenRouter), and it runs no model itself.

Features

• Agents that remember. Each Agent keeps its memory as Markdown files under git. The Learning Feed shows each memory change, with a revert.
• A computer for each Agent. You watch its screen live, take control at any time and give control back.
• Connections with scoped access. Mail and calendar through Google, a Grant for each Agent, approvals for actions that change the world, and a Vault whose secrets the model never sees.
• Proactive work. Schedules and Event Subscriptions start work while you are away.
• A phone number and a mailbox. An Agent can hold its own number, make and answer calls, and send and receive mail from its own address.
• Software and Plugins. Agents build versioned tools that other Agents use and fork. An Administrator installs Plugins: Skills and MCP servers.
• Hosts. An Agent runs an approved command on your own machine through the Client App.
• One person or a team. One installation is one Org. Each Person has a private Workspace, and an Administrator manages the People, the spend and the providers.

Quick start

A release installs in two ways:

• The Client App (macOS and Linux) runs Pagis on your own computer or connects to a server. Download it from the latest release, and the Quickstart installs it.
• The Headless Server (Linux VM with Docker) runs Pagis for a team from the image ghcr.io/pagis-co/pagis-server:<release>. Deploy with Compose starts it.

The release holds one Client App package for each platform. <release> is the release number, such as 0.1.0.

On Ubuntu 24.04 and later, use the deb. The release key signs the checksum list of the Linux packages, and Verify a download tells how to check it.

From source

You need Rust, Node.js (the version in .nvmrc) and, for the Agents' computers, Docker.

(cd ui && npm ci && npm run build)
cargo run -p pagis -- --local

--local makes a local installation. At the first run, the daemon prints a one-time sign-in link and opens the browser on it. pagis --help lists every flag and command.

Analytics

A release build sends anonymous analytics to PostHog once a day: the release, the operating system, the kind of installation, the number of People and Agents as ranges, and which features are in use. The events carry a random installation ID. They hold no content, no names, no addresses and no IP address. An Administrator turns them off in System Settings in the Administration Interface, and DO_NOT_TRACK=1 in the environment of the daemon stops them too. A build from source sends nothing. ADR-0026 holds the details.

Documentation

The user documentation is at docs.pagis.co, and docs-site/ holds its source. Each release deploys the pages of its own tree, and a manual run of the Docs workflow deploys main. The documents below are for the people who build Pagis.

Contributing

CONTRIBUTING.md describes the development setup, the checks and how a pull request merges.

Security

Report a vulnerability through GitHub private vulnerability reporting. Do not open a public issue for it. SECURITY.md holds the policy.

License

Pagis is licensed under the MIT License. Each server package also carries LICENSE.gog and THIRD_PARTY_NOTICES for the third-party programs that it ships.