Skip to main content
This tutorial takes you from installation to a working mesh in five steps. What you are building is a Monad Mesh: a durable agent team the daemon owns. It starts with one member — Monad Agent Runtime, Monad’s bundled first-party agent — and grows by adding other agent runtimes without moving any work. Monad stores its own state on your machine. Requests to the model provider you choose still leave the machine. 简体中文版：开始使用 Monad

​
1. Install Monad

Use one of the installation methods in the root README. In an interactive terminal, the installer starts the daemon and opens the Web UI on macOS, Linux, and Windows, including the setup flow for a new installation. After an automated or quiet install, run monad up to do the same. The mesh owns the team and its work. Closing a browser or terminal does not transfer that authority to another client.

​
2. Connect a model

Your first mesh member is Monad Agent Runtime, and it needs a model provider. Monad does not bundle local inference. If you plan to work only through third-party agent providers such as Codex or Claude Code, they authenticate separately — skip to step 5. Choose one setup route: The setup flow tests the connection before saving it and lets you choose a default model. See model providers for provider-specific configuration. Check the result:

​
3. Run your first session

A session is one durable work thread. It survives client closure and daemon restart, and you can operate it from any client. Run a one-off task:
Create and observe a named session:
The same session appears in the Web UI because state lives in the daemon. See sessions and the CLI reference for branching, restoration, and scripting.

​
4. Handle an approval

When an agent requests a gated action, the turn stops and waits for a decision. The approval shows the requested action and its consequence. Keep these rules in mind:
  • Denial is a normal result: the agent receives it and can continue
  • The gate fails closed: a high-risk call is denied when no client can answer
  • Sandboxing is separate: approval does not replace process and network containment
See sandbox backends before increasing agent autonomy.

​
5. Grow the mesh

Add members first, then capability. Existing sessions and history stay where they are. Then extend what a member can do: Type / in the chat input to inspect available skills and commands.

​
Continue from here