Anthropic Code Interpreter - Dartantic AI Documentation

Anthropic’s code interpreter runs Python in a sandboxed environment. State persists across turns, and generated files are automatically downloaded.

Enable Code Interpreter

final agent = Agent(
  'anthropic',
  chatModelOptions: const AnthropicChatOptions(
    serverSideTools: {AnthropicServerSideTool.codeInterpreter},
  ),
);

Multi-Turn Execution

State persists across conversation turns:

final history = <ChatMessage>[];

// Step 1: Calculate data
await for (final chunk in agent.sendStream(
  'Calculate the first 10 Fibonacci numbers and store them in a variable '
  'called "fib_sequence". Create a CSV file called "fibonacci.csv".',
)) {
  stdout.write(chunk.output);
  history.addAll(chunk.messages);
}

// Step 2: Use the data from step 1
await for (final chunk in agent.sendStream(
  'Using fib_sequence, create a plot showing the golden ratio convergence. '
  'Save it as "golden_ratio.png".',
  history: history,
)) {
  stdout.write(chunk.output);
  history.addAll(chunk.messages);
}

Generated Files

Files created during code execution are attached as DataPart instances:

for (final msg in history) {
  for (final part in msg.parts) {
    if (part is DataPart) {
      print('Generated: ${part.name} (${part.mimeType})');
      File('output/${part.name}').writeAsBytesSync(part.bytes);
    }
  }
}

Anthropic’s code interpreter supports common Python libraries including matplotlib for plotting, pandas for data manipulation, and reportlab for PDF generation.

Streaming Code Events

Code execution events stream through metadata['code_interpreter']:

await for (final chunk in agent.sendStream(prompt)) {
  final events = chunk.metadata['code_interpreter'] as List?;
  if (events != null) {
    for (final event in events) {
      if (event['type'] == 'code_delta') {
        stdout.write(event['delta']);
      }
    }
  }
}

Example

Anthropic Code Interpreter demo