# Build a Coding Agent Using OpenCode SDK and Daytona (/docs/guides/opencode/opencode-sdk-agent)

This guide walks you through running the [OpenCode](https://opencode.ai/) autonomous coding agent inside a secure Daytona sandbox. The OpenCode server runs fully isolated in the sandbox and the script connects to the server using the [OpenCode SDK](https://www.npmjs.com/package/@opencode-ai/sdk).

The agent can develop full-stack web apps, write code in any language, install dependencies, and run scripts. It can also start and manage dev servers, and generate [preview links](/docs/preview/) for live apps.

### 1. Workflow Overview [#1-workflow-overview]

When you run the script, it creates a Daytona sandbox, installs OpenCode, and starts the OpenCode server inside the sandbox. You send prompts and see responses in your terminal. On exit, the sandbox is deleted.

```
$ npm run start
Creating sandbox...
Installing OpenCode in sandbox...
Preview: https://4096-0626a08c-08e7-4fbe-9799-d160bac66b08.daytonaproxy01.net
Press Ctrl+C at any time to exit.
User: Create a minimalist collaborative whiteboard app
Thinking...
🔨 ✓ Run: Check current directory structure
✓ 3 todos
📝 Add home/daytona/SPEC.md
✓ 2 todos
📝 Add home/daytona/index.html
✓ 1 todos
✓ home/daytona/SPEC.md
✓ https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400;500&family=Outfit:wght@400;500;700&display=swap (text/css; charset=utf-8)
✓ home/daytona/index.html
🔨 ✓ Run: List files in directory
🔨 ✓ Run: Start HTTP server for whiteboard
🔨 ✓ Run: Check if server is running
✓ 0 todos
Your minimalist collaborative whiteboard app is ready!

**Preview URL:** https://8080-0626a08c-08e7-4fbe-9799-d160bac66b08.daytonaproxy01.net

## SyncBoard - Features

**Drawing Tools:**
- Pencil (freehand drawing)
- Line, Rectangle, Ellipse (shape tools with preview)
- Eraser (removes intersecting strokes)

**Customization:**
- 6 vibrant color presets
- 3 stroke widths (thin/medium/thick)
- Keyboard shortcuts: `P` `L` `R` `E` `X` for tools, `1` `2` `3` for stroke widths

**Collaboration:**
- Simulated real-time collaboration with other users
- Unique user colors for each participant
- Auto-saves to localStorage every 2 seconds

**Actions:**
- Undo/Redo (stores 20 actions, Ctrl+Z / Ctrl+Shift+Z)
- Clear canvas (with confirmation dialog)
- Coordinates display in status bar

**Design:**
- Dark theme with coral/teal accents
- Dot grid background pattern
- Smooth animations and hover effects
- Responsive layout for mobile devices

User: 
Cleaning up...
```

When your task involves running or previewing a web application, the agent can host the app and generate a [preview link](/docs/preview/) for you to inspect the live result:

<Image
  src="/docs/images/opencode-sdk-syncboard.gif"
  alt="SyncBoard collaborative whiteboard app generated by OpenCode agent"
  width={600}
/>

You can continue interacting with your agent until you are finished. When you exit the program, the sandbox will be deleted automatically.

### 2. Project Setup [#2-project-setup]

#### Clone the Repository [#clone-the-repository]

First, clone the Daytona [repository](https://github.com/daytona/guides.git) and navigate to the OpenCode SDK example:

```bash
git clone https://github.com/daytona/guides.git
cd guides/typescript/opencode/opencode-sdk
```

#### Configure Environment [#configure-environment]

Get your API key from the [Daytona Dashboard](https://app.daytona.com/dashboard/keys).

Copy `.env.example` to `.env` and add your key:

```bash
DAYTONA_API_KEY=your_daytona_key
```

#### Local Usage [#local-usage]

<Callout type="note" title="Node.js Version">
Node.js 18 or newer is required to run this example. Please ensure your environment meets this requirement before proceeding.
</Callout>

Install dependencies and run the example:

```bash
npm install
npm run start
```

The agent will start and wait for your prompt.

### 3. Understanding the Script [#3-understanding-the-script]

OpenCode uses a client-server model: The [server](https://opencode.ai/docs/server/) manages coding agents, sessions and configuration. Clients communicate with the server [over HTTP](https://opencode.ai/docs/server/#apis) to run prompts in sessions and receive streamed responses.

#### Initialization [#initialization]

On startup, the script:

1. Creates a new [Daytona sandbox](/docs/sandboxes).
2. Installs OpenCode globally in the sandbox.
3. Starts the OpenCode server in the sandbox.
4. Gets the preview URL for the server and creates a client with that base URL.
5. Enters the readline loop to send prompts to the server and receive streamed responses.
6. On Ctrl+C, deletes the sandbox and exits.

#### OpenCode Server [#opencode-server]

The OpenCode server runs inside the sandbox and handles all agent work: LLM calls, tools (bash, write, etc.), and code execution. The script starts it with:

```ts
const envVar = injectEnvVar('OPENCODE_CONFIG_CONTENT', opencodeConfig)
const command = await sandbox.process.executeSessionCommand(sessionId, {
  command: `${envVar} opencode serve --port ${PORT} --hostname ${HOSTNAME}`,
  runAsync: true,
})
```

A custom configuration is injected which includes a system prompt. The system prompt tells the agent it runs in a sandbox, to use `/home/daytona` for file operations, and the preview URL pattern so it can provide correct preview links.

#### OpenCode Client [#opencode-client]

The client runs on your machine and talks to the server using the [OpenCode SDK](https://opencode.ai/docs/sdk/). `Session.create` sets up the SDK client and subscribes to the event stream so the client can show tool activity (e.g. “✓ Run: …”, “📝 Add …”) as the agent works:

```ts
const client = createOpencodeClient({ baseUrl })
const sessionRes = await client.session.create({ body: { title: 'Daytona query' } })
const sessionId = sessionRes.data?.id
if (!sessionId) throw new Error('Failed to create OpenCode session:' + sessionRes.error)
const events = await client.event.subscribe()
```

To send a prompt, the client calls `session.prompt` and processes events from the existing stream to show tool activity:

```ts
const promptPromise = this.client.session.prompt({
  path: { id: this.sessionId },
  body: { parts: [{ type: 'text', text: query } satisfies TextPartInput] },
})

for await (const event of takeUntil(this.events.stream, promptPromise)) {
  printEvent(this.sessionId, event)
}
```

#### Main loop [#main-loop]

When the server is ready, the script creates a session and runs a readline loop:

```ts
const session = await Session.create(baseUrl)
const rl = readline.createInterface({ input: process.stdin, output: process.stdout })

while (true) {
  const query = await new Promise<string>((resolve) => rl.question('User: ', resolve))
  if (!query.trim()) continue
  await session.runQuery(query)
}
```

The readline loop waits for user input, sends it to the agent, and prints the response.

**Key advantages:**

- Secure, isolated execution in Daytona sandboxes
- OpenCode SDK client in your terminal; server runs in the sandbox
- Support for 75+ LLM providers
- All agent code execution happens inside the sandbox
- Automatic preview link generation for deployed services
- Custom agent configuration for Daytona-specific workflows