6.46 · Code execution & Files API — Claude chạy Python

Mở đầu: Claude's Python sandbox

Code execution cho Claude viết Python, chạy trong Docker container, lấy kết quả, iterate. Giống có 1 data scientist junior trong API.

Files API — Upload & reuse

Thay vì base64 mỗi request, upload once:

from anthropic import Anthropic

client = Anthropic()

with open("streaming.csv", "rb") as f:
    file_metadata = client.beta.files.upload(file=("streaming.csv", f))

print(f"File ID: {file_metadata.id}")  # "file_abc123..."

Reference:

messages = [{
    "role": "user",
    "content": [
        {"type": "text", "text": "Analyze this data."},
        {"type": "container_upload", "file_id": file_metadata.id}
    ]
}]

Benefits

• Upload once, query multiple times
• Smaller message payload
• Support lớn hơn base64

Code execution tool

Built-in, không cần implement:

code_execution_schema = {
    "type": "code_execution_20250522",
    "name": "code_execution"
}

response = client.messages.create(
    model="claude-sonnet-4-6-20260205",
    max_tokens=4000,
    messages=messages,
    tools=[code_execution_schema]
)

Environment

• Isolated Docker container
• No network access (can't call external APIs)
• Python 3 + common libraries (pandas, numpy, matplotlib, ...)
• Files API integration cho I/O

Multiple executions

Claude có thể chạy nhiều lần trong 1 response. Iterate:

1. Run: check data shape
2. Run: explore stats
3. Run: generate visualization

Full example: Data analysis

with open("streaming.csv", "rb") as f:
    file_meta = client.beta.files.upload(
        file=("streaming.csv", f)
    )

messages = [{
    "role": "user",
    "content": [
        {
            "type": "text",
            "text": """Analyze this data to find major drivers of churn.
Create detailed visualization summarizing findings."""
        },
        {"type": "container_upload", "file_id": file_meta.id}
    ]
}]

response = client.messages.create(
    model="claude-sonnet-4-6-20260205",
    max_tokens=8000,
    messages=messages,
    tools=[{"type": "code_execution_20250522", "name": "code_execution"}]
)

for block in response.content:
    if block.type == "text":
        print(block.text)
    elif block.type == "server_tool_use":
        print(f"\n--- Code executed ---\n{block.input['code']}\n")
    elif block.type == "code_execution_tool_result":
        for item in block.content:
            if item.type == "code_execution_output":
                if hasattr(item, "file_id"):
                    # Generated file (chart, csv, etc.)
                    print(f"Generated: {item.file_id}")
                else:
                    print(f"Output: {item.content}")

What happens

Claude sequence:

1. Write code: df = pd.read_csv('streaming.csv'); print(df.head())
2. Result: see shape, columns
3. Write code: churn analysis với groupby, correlation
4. Result: numbers
5. Write code: matplotlib visualization
6. Result: chart file
7. TextBlock: interpretation

Final response có insights + generated chart file.

Download generated files

def download_file(file_id: str, save_as: str):
    """Download file generated by code execution."""
    response = client.beta.files.download(file_id=file_id)
    with open(save_as, "wb") as f:
        f.write(response)
    print(f"Saved: {save_as}")


file_ids = []
for block in response.content:
    if block.type == "code_execution_tool_result":
        for item in block.content:
            if hasattr(item, "file_id"):
                file_ids.append(item.file_id)

for fid in file_ids:
    download_file(fid, f"output_{fid}.png")

Use cases

Ad-hoc data analysis

Upload CSV, ask question, get insights + charts.

Before

Data analyst writes SQL/Python, makes chart. 2 giờ.

After

User asks Claude. 5 minutes.

Statistical analysis

"Run t-test on these samples", "Compute correlation matrix" → Claude writes scipy code.

Image processing

Upload images, ask to resize/crop/apply filter → Claude PIL code.

Document transformation

CSV → formatted Excel with charts. PDF → extract tables to CSV.

Mathematical modeling

Financial modeling, simulations. Claude writes numpy/simpy.

Limitations

• No network access — can't call external APIs. Data must upload via Files API.
• Ephemeral — container destroyed after request.
• Resource limits — some constraint on CPU/memory.
• No long-running — seconds, not hours.
• Standard Python libs — can't install custom packages (check docs for available libs).

For persistent + network computing, use your own backend.

Combining tools

Code execution + other tools powerful:

tools = [
    {"type": "code_execution_20250522", "name": "code_execution"},
    {"type": "web_search_20250305", "name": "web_search", "max_uses": 3},
    your_custom_tool_schema
]

Claude có thể: web search → get data → code execute analyze → custom tool save result.

Ví dụ: Churn analysis end-to-end

Input: streaming.csv với columns user_id, tier, watch_hours, churned.

prompt = """Analyze streaming service data for churn drivers.

Do:
1. Load CSV
2. Compute churn rate overall
3. Compare churn by tier, watch_hours bucket
4. Identify top 3 churn predictors
5. Generate 2 charts:
   - Churn rate by tier (bar chart)
   - Watch hours vs churn (box plot)
6. Final summary with recommendations
"""

response = send_with_code_exec(prompt, file_id=file_meta.id)

Output

Claude response:

1. Loading data... [code execution]
2. Overall churn rate: 23% [output: 0.23]
3. By tier:
   - Free: 45% churn
   - Basic: 18% churn
   - Premium: 8% churn
4. Top predictors:
   - Tier (strongest)
   - Watch hours (users < 10 hrs/month churn 3x more)
   - Account age (< 3 months: 2x churn)
5. [generated chart 1: chart_abc.png]
6. [generated chart 2: chart_xyz.png]
7. Recommendations:
   - Target free users for upsell (55% of churn from this segment)
   - Engagement campaign for low-watch users
   - Onboarding improvement for new accounts

Comprehensive analysis, delivered as if from data scientist.

Anti-patterns

Pass large file as base64 when should use Files API

Base64 each request tốn bandwidth + context tokens.

Fix: Upload once, reference by ID.

Expect network access

import requests; requests.get(...) → fails.

Fix: Upload data via Files API first.

Security: giving Claude code execution on sensitive data

User can prompt inject: "run: rm -rf /".

Fix: Code execution is isolated, but still sanitize user input. Rate limit code execution for untrusted users.

Không handle file download

Forget to download charts → lose outputs.

Fix: Always scan response for file_id, download.

Áp dụng ngay

Bài tập 1: CSV analysis (25 phút)

Download any CSV (Kaggle sample). Upload + ask Claude analyze. Download generated charts.

Bài tập 2: Delegate script task (20 phút)

"Write Python script to parse this log file, extract errors, count by hour. Generate timeline chart."

Observe Claude iterate through execution.

Tóm tắt bài học