---
title: Use Mastra Coding Agent with Daytona
description: Learn how to integrate Daytona sandboxes with the Mastra coding agent for secure, isolated AI-powered development.
guide:
  tags: ["Mastra"]
sidebar:
  label: "Mastra"
---

This guide demonstrates how to configure the [Mastra coding agent](https://github.com/mastra-ai/template-coding-agent) to use Daytona sandboxes, enabling you to leverage AI capabilities for any coding-related task in a secure, isolated environment.

---

### 1. Workflow Overview

Once configured, you can use Mastra Studio to interact with the coding agent through a ChatGPT-like interface. This enables human-in-the-loop workflows where you can guide the agent, review its decisions, and iterate on solutions in real-time; all while the agent executes tasks securely within Daytona sandboxes.

### 2. Project Setup

<Callout type="note" title="Node.js Version">
Node.js version 20 or higher is required to run the coding agent. Please ensure your environment meets this requirement before proceeding.
</Callout>

#### Clone the Repository

Clone the Mastra coding agent template repository, which includes the agent implementation and Daytona integration:

```bash
git clone https://github.com/mastra-ai/template-coding-agent.git
cd template-coding-agent
```

#### Configure Environment

Create a `.env` file in the project root directory:

```bash
touch .env
```

The `.env` file requires the following configuration:

- **LLM Provider**: The AI model provider for your coding agent
- **Model**: The specific model to use from your chosen provider
- **Sandbox Provider**: Daytona configuration for isolated execution

For this guide, we'll use OpenAI as the LLM provider with the `gpt-4o-mini` model:

```bash
OPENAI_API_KEY=your_openai_key
MODEL=openai/gpt-4o-mini
```

The agent works with any provider supported by [Mastra's model router](https://mastra.ai/models) - set `MODEL` in `provider/model` format and the matching provider API key. For example, to use [Fireworks AI](https://fireworks.ai/):

```bash
FIREWORKS_API_KEY=your_fireworks_key
MODEL=fireworks-ai/accounts/fireworks/models/kimi-k2p6
```

Next, configure Daytona as your sandbox provider by adding your API key (available from the [Daytona Dashboard](https://app.daytona.io/dashboard/keys)):

```bash
DAYTONA_API_KEY=your-daytona-api-key-here
```

#### Install Dependencies

Install the required packages using pnpm:

```bash
pnpm install
```

### 3. Running the Agent

Mastra Studio provides a ChatGPT-like interface for interacting with your coding agent. This mode offers:

- **Conversation History**: Previous conversations are stored and organized in threads
- **Visual Debugging**: Inspect agent execution steps, workflow, and tool calls
- **Model Switching**: Easily switch between different AI models
- **Tool Inspection**: View which tools your agent is using in real-time

For a complete overview of all features and capabilities, visit the [Mastra Studio documentation](https://mastra.ai/docs/getting-started/studio).

Start the dev server with:

```bash
pnpm run dev
```

If the dev server starts successfully, you'll see the terminal output displaying the URLs where you can access Mastra Studio:

```bash
│ Studio: http://localhost:4111
│ API:    http://localhost:4111/api
```

Once started, open the provided URL in your browser to access the interactive interface. You can interact with your agent while monitoring its workflow in the terminal, where detailed logs show execution steps and low-level parameters.

Below is an example of terminal logs generated when the agent calls the `writeFile` tool to create a JavaScript file with a basic "Hello, world!" output:

```json
{
  "text": "",
  "toolCalls": [
    {
      "type": "tool-call",
      "runId": "ab2a1d08-91c6-4028-9046-3446a721527f",
      "from": "AGENT",
      "payload": {
        "toolCallId": "call_NiLLgBmgrYLSL0MsrG54E4A5",
        "toolName": "writeFile",
        "args": {
          "sandboxId": "2152d23b-5742-47c2-9992-4414d4144869",
          "path": "hello.js",
          "content": "console.log('Hello, world!');"
        },
        "providerMetadata": {
          "openai": {
            "itemId": "fc_00bba3412cd22a2b0069399fbaeef881909b0583f359cbc33c"
          }
        }
      }
    }
  ],
  "toolResults": [
    {
      "type": "tool-result",
      "runId": "ab2a1d08-91c6-4028-9046-3446a721527f",
      "from": "AGENT",
      "payload": {
        "args": {
          "sandboxId": "2152d23b-5742-47c2-9992-4414d4144869",
          "path": "hello.js",
          "content": "console.log('Hello, world!');"
        },
        "toolCallId": "call_NiLLgBmgrYLSL0MsrG54E4A5",
        "toolName": "writeFile",
        "result": {
          "success": true,
          "path": "/home/daytona/hello.js"
        },
        "providerMetadata": {
          "openai": {
            "itemId": "fc_00bba3412cd22a2b0069399fbaeef881909b0583f359cbc33c"
          }
        }
      }
    }
  ],
  "finishReason": "tool-calls",
  "usage": {
    "inputTokens": 4243,
    "outputTokens": 53,
    "totalTokens": 4296,
    "reasoningTokens": 0,
    "cachedInputTokens": 4096
  },
  "runId": "ab2a1d08-91c6-4028-9046-3446a721527f"
}
```

These logs show complete visibility into tool execution including arguments, results, token usage with caching metrics, and unique identifiers for debugging.

### 4. Example Usage

Let's walk through a practical example to see the Mastra coding agent and Daytona sandbox integration in action. We'll ask the agent to complete the following task:

> _"Create a Python file with a function that reverses a string, write a test file for it, and run the tests."_

#### Agent Workflow

The agent executes this task through a series of coordinated steps, utilizing different tools to interact with the Daytona sandbox:

##### 1. Create Sandbox

The agent first provisions a new Daytona sandbox for Python development:

**Tool:** `createSandbox`

**Arguments:**

```json
{
  "name": "reverse_string_project",
  "labels": null,
  "language": "python",
  "envVars": null
}
```

**Result:**

```json
{
  "sandboxId": "bdfa9456-4945-43a7-97df-b9bfbcbdc62e"
}
```

##### 2. Write Project Files

Next, the agent creates both the implementation and test files in a single operation:

**Tool:** `writeFiles`

**Arguments:**

```json
{
  "sandboxId": "bdfa9456-4945-43a7-97df-b9bfbcbdc62e",
  "files": [
    {
      "path": "reverse_string.py",
      "data": "def reverse_string(s):\n    \"\"\"Return the reversed version of the input string s.\"\"\"\n    return s[::-1]\n"
    },
    {
      "path": "test_reverse_string.py",
      "data": "import unittest\nfrom reverse_string import reverse_string\n\nclass TestReverseString(unittest.TestCase):\n    def test_regular(self):\n        self.assertEqual(reverse_string(\"hello\"), \"olleh\")\n    def test_empty(self):\n        self.assertEqual(reverse_string(\"\"), \"\")\n    def test_single_char(self):\n        self.assertEqual(reverse_string(\"a\"), \"a\")\n    def test_numbers(self):\n        self.assertEqual(reverse_string(\"12345\"), \"54321\")\n\nif __name__ == \"__main__\":\n    unittest.main()\n"
    }
  ]
}
```

**Result:**

```json
{
  "success": true,
  "filesWritten": [
    "/home/daytona/reverse_string.py",
    "/home/daytona/test_reverse_string.py"
  ]
}
```

##### 3. Execute Tests

Finally, the agent runs the test suite to verify the implementation:

**Tool:** `runCommand`

**Arguments:**

```json
{
  "sandboxId": "bdfa9456-4945-43a7-97df-b9bfbcbdc62e",
  "command": "python3 test_reverse_string.py",
  "envs": null,
  "workingDirectory": null,
  "timeoutSeconds": 20,
  "captureOutput": true
}
```

**Result:**

```json
{
  "success": true,
  "exitCode": 0,
  "stdout": "....\n----------------------------------------------------------------------\nRan 4 tests in 0.000s\n\nOK\n",
  "command": "python3 test_reverse_string.py",
  "executionTime": 218
}
```

#### Agent Response

After successfully completing all steps, the agent provides a detailed summary:

```
The project is complete! Here's what was done:

✓ Created reverse_string.py with a function to reverse a string.
✓ Created test_reverse_string.py using Python's unittest to test the function with several cases.
✓ Ran the tests—all tests passed successfully.

Test Output:
....
----------------------------------------------------------------------
Ran 4 tests in 0.000s

OK

Would you like to see the code for either file or make any changes?
```

This example demonstrates how the agent autonomously handles the entire development workflow, from environment setup to test execution, all within the secure Daytona sandbox environment.

<Callout type="tip" title="Sandbox Management">
After the agent completes its task, you'll see the active sandbox in your [Daytona Dashboard](https://app.daytona.io/dashboard). It's recommended to clean up resources when finished, unless your use case requires the sandbox to remain active (e.g. for preview URLs or ongoing development).
</Callout>

**Key Advantages**

- **Secure Isolation**: All agent operations run in isolated Daytona sandboxes, protecting your local environment
- **Multi-Language Support**: Execute code across different programming languages without local setup
- **Enhanced Debugging**: Use Mastra Studio to visualize and debug agent workflows in real-time
- **Scalable Execution**: Leverage Daytona's cloud infrastructure for resource-intensive tasks