# Deploy an agent

> Deploy a Nested agent: choose a base, packs, and a coding agent (Cursor, Claude Code, Kimi CLI), attach secrets, and choose who can chat with it.

Section: Your agents · Updated: 2026-11-19 · Canonical: https://nested.me/docs/deploy

An agent in Nested is a deployment: a running `cloud-agent` in your team’s Kubernetes cluster, built from a base image, extended with packs, and driven by a coding agent such as Claude Code or Cursor. The **Deploy a focused agent** wizard asks only what an agent needs to start. Cluster placement and Helm overrides stay with operators in the console.

## Before you start

- Sign in and pick a team. An agent always belongs to one team.
- Know which coding agent you want. Cursor needs a `CURSOR_API_KEY`; Claude Code and Kimi CLI can sign in after the agent starts.
- Optional: save API keys ahead of time in **Settings → Secrets**. You can also create them inside the wizard.

## Open the wizard

Use any of these entry points:

- **New agent** at the end of the **Agents** group in the sidebar. On desktop it opens the wizard as a dialog.
- **New agent** on the home screen when you have no agents yet.
- **New agent** in the command palette (Cmd/Ctrl+K).
- **Settings → Agents → New agent** (the usual path on a phone).

## Steps

1. **Choose.** Open an agent that already answers anyone (marked **Answers anyone**), or select **Create my own agent**. Opening a shared agent takes you straight to chat; the rest of the steps apply only when you create your own.
2. **Name.** Enter an **Agent name**: 3–63 lowercase letters, numbers, or dashes, starting with a letter, unique in the team. On the same step, answer **Who can chat with this agent?** with **Only me** or **Anyone signed in**.
3. **Base.** Pick the **Base agent**, the Nested catalog image the deployment is built from. Nested skips this step when the catalog has only one base.
4. **Pack.** Select zero or more **Packs**. Packs add skills, formulas, MCP servers, and tools; the agent uses each pack’s latest revision. Select **Explore** to preview a pack’s contents before you add it. You can continue without a pack.
5. **ACP.** Choose the **ACP agent**, the coding agent that runs inside the deployment (see the table below).
6. **Secrets.** Attach a saved secret to each environment variable the packs and the ACP agent declare, or select **Create a new secret** and **Save and attach**. Each secret is injected under the variable name you set, not under the secret’s own name.
7. **Confirm.** Review the summary. If your team’s cluster has Kata installed, choose the **Runtime** here (see the table below). Select **Deploy**. Nested deploys the agent onto your team’s cluster and opens it in chat when it is running.

### Runtime

The **Runtime** choice appears only when the cluster offers Kata. Otherwise every agent uses the standard runtime.

| Option | What it means |
| --- | --- |
| **Standard** (default) | Shared kernel, fastest start. Right for most agents. |
| **Isolated VM** | The agent runs in its own virtual machine. Stronger isolation, but it starts slower, builds run about 1.6× slower, and it reserves its full memory limit. |

> [!WARNING]
> You can’t change the runtime after the agent is created. To move an agent to another runtime, create a new agent.

### ACP agents

| Option | How it signs in | Notes |
| --- | --- | --- |
| **Cursor** | API key | Needs a `CURSOR_API_KEY` secret. Ready immediately on boot. |
| **Claude Code** | Web sign-in | No key during setup. You sign in to Claude in chat after the agent starts. |
| **Kimi CLI** | Device sign-in | Signs in with a device code in chat, or attach a `KIMI_API_KEY` secret to skip it. |
| **Choose from registry** | Varies | Community and third-party ACP agents from the registry. |
| **Manual entry** | Varies | Runs any ACP-compatible command already in the base image, with optional arguments and environment variables. |

> [!TIP]
> Task acceptance and secret visibility are different settings. **Only me / Anyone signed in** on the Name step controls who can chat with the agent. **Only me / Everyone** on a secret controls who can attach that secret.

## Example

Create `research`, keep **Only me**, add a browsing pack, choose **Claude Code**, and attach a saved `SEARCH_API_KEY` secret. After **Deploy**, the agent appears in the sidebar **Agents** group. When it is running, a sign-in card asks you to sign in to Claude once; then start your first task.

## Questions

### Which coding agents can a Nested agent run?

The wizard offers Cursor, Claude Code, and Kimi CLI, plus any ACP agent from the registry or a manual command that is installed in the base image.

### Can I deploy an agent without a pack?

Yes. Packs are optional on the **Pack** step; you can attach them later in agent settings → **Packs**.

### Who can use my agent after I deploy it?

With **Only me**, only you can start tasks on it. With **Anyone signed in**, everyone in the team sees it in their picker. Only you can change, stop, or delete it.

### Can a Nested agent run in its own virtual machine?

Yes, when your cluster has Kata installed. Choose **Isolated VM** as the **Runtime** on the wizard’s Confirm step. The choice is fixed once the agent is created.

[Open in Nested](https://app.nested.me/agents/new)

## Related

- [Agent settings](https://nested.me/docs/agent-settings)
- [Secrets](https://nested.me/docs/secrets)
- [Agent sign-in, accounts, and auto-stop](https://nested.me/docs/agent-accounts)
- [Packs](https://nested.me/docs/packs-page)
