latestaiagents/code-execution

Code Execution Tool

The Code Execution tool runs Python in an Anthropic-hosted sandbox as part of a model response. Use it when you need the model to actually compute, not just describe.

When to Use

• Data analysis and transformation (CSV, JSON, parquet)
• Chart and plot generation (matplotlib, plotly)
• Math verification — let the model check its own work
• Running unit tests the model just wrote
• Any task where "the answer is whatever the code prints"

Enabling

const response = await client.beta.messages.create(
  {
    model: "claude-sonnet-4-6",
    max_tokens: 4096,
    tools: [{ type: "code_execution_20250522", name: "code_execution" }],
    messages: [{ role: "user", content: "Analyze the attached CSV and plot monthly revenue." }],
  },
  { headers: { "anthropic-beta": "code-execution-2025-05-22" } },
);

No tool-use loop to manage — execution happens server-side and results flow back in the response.

What's in the Sandbox

• Python 3.12+
• Preinstalled: pandas, numpy, matplotlib, scipy, scikit-learn, pillow, requests, beautifulsoup4, and more
• Ephemeral filesystem (cleared between calls unless using container persistence — see below)
• No network access by default (some betas allow it)
• CPU + RAM limits (conservative — don't train models here)

File Upload

Upload files with the Files API, then reference them in the message:

const file = await client.beta.files.upload({ file: fs.createReadStream("sales.csv") });

const response = await client.beta.messages.create(
  {
    model: "claude-sonnet-4-6",
    max_tokens: 4096,
    tools: [{ type: "code_execution_20250522", name: "code_execution" }],
    messages: [
      {
        role: "user",
        content: [
          { type: "container_upload", file_id: file.id },
          { type: "text", text: "Compute YoY growth from sales.csv." },
        ],
      },
    ],
  },
  { headers: { "anthropic-beta": "code-execution-2025-05-22,files-api-2025-04-14" } },
);

The file appears at /mnt/user-data/ inside the sandbox.

Reading Execution Results

for (const block of response.content) {
  if (block.type === "code_execution_tool_result") {
    const result = block.content;
    console.log("stdout:", result.stdout);
    console.log("stderr:", result.stderr);
    if (result.return_code !== 0) console.log("FAILED");
    for (const file of result.files ?? []) {
      // file is a generated image/data file with file_id; download via Files API
    }
  }
}

Container Persistence

Re-use the sandbox across calls by passing the container ID:

// First call creates a container
const first = await client.beta.messages.create({ /* ... */ });
const containerId = first.container?.id;

// Second call re-uses it — pip installs, written files, variables all persist
const second = await client.beta.messages.create({
  /* ... */
  container: containerId,
  messages: [
    ...firstMessages,
    { role: "user", content: "Now compute the 7-day rolling avg on that same DataFrame." },
  ],
});

Containers auto-expire after inactivity. Use persistence for multi-turn data analysis; skip it for one-shot.

Chart Output

matplotlib figures are returned as image files the Files API can serve:

import matplotlib.pyplot as plt
plt.plot([1,2,3], [4,5,6])
plt.savefig("/tmp/out.png")

The tool result includes a file_id. Download and render:

const bytes = await client.beta.files.content(fileId);
await fs.writeFile("out.png", bytes);

Pairing with Extended Thinking

For analysis tasks, enable thinking so the model plans before coding:

const response = await client.beta.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 16_000,
  thinking: { type: "enabled", budget_tokens: 8000 },
  tools: [{ type: "code_execution_20250522", name: "code_execution" }],
  messages: [...],
});

Enable interleaved thinking (see the extended-thinking skill) so it can reflect on code output before the next step.

Limitations

• No persistent installed packages across new containers — use pip install at start
• No unrestricted network by default — fetch external data beforehand
• File system wiped on container expiry
• Execution time-limited per call (typically 60-120s)
• Not suitable for long-running training or web scraping

Security

The sandbox is isolated — user data doesn't leak between containers. Your risk is that the model might:

• Write your uploaded data to logs you later read back
• Print secrets included in messages to stdout

Scrub outputs before persisting (see mcp-security-sandboxing for redaction patterns).

Anti-Patterns

1. Using it for code writing when no execution is needed — waste of server resources
2. Uploading huge files (> 100 MB) — slow and often unnecessary; sample first
3. Not persisting the container for multi-turn analysis — pays setup cost each time
4. Ignoring stderr — silent failures yield wrong answers

Best Practices

1. Enable for analysis, math, tests, plots. Don't enable for pure text tasks
2. Use container persistence for multi-turn analysis
3. Pair with extended thinking for complex analysis
4. Always check return_code — model can confidently report results from failed code
5. Retrieve generated plots via Files API; don't re-ask the model to describe them
6. Limit file upload size; pre-sample large datasets client-side