Cloud Sandboxes

Sandboxes are isolated cloud execution environments that give your agents the ability to run code, manage files, and execute processes. They boot in ~60ms and provide a full Linux environment.

Why Sandboxes?

LLMs can reason about code, but they can't execute it. Sandboxes bridge this gap:

Without Sandbox	With Sandbox
Agent can only suggest code	Agent can run and verify code
No file manipulation	Full filesystem access
No web browsing	Browser automation available
Stateless	Persistent state across interactions

Quick Start

import requests

BASE = "https://api.sandbase.ai"
HEADERS = {"Authorization": "Bearer sk-sb-YOUR_KEY"}

# 1. Create a sandbox (boots in ~60ms)
sandbox = requests.post(f"{BASE}/sandboxes", headers=HEADERS, json={
    "templateID": "code_interpreter",
    "timeout": 300
}).json()

sandbox_id = sandbox["sandboxID"]

# 2. Execute code
result = requests.post(f"{BASE}/sandboxes/{sandbox_id}/processes", headers=HEADERS, json={
    "cmd": "python3",
    "args": ["-c", "import pandas as pd; print(pd.__version__)"]
}).json()

print(result["stdout"])  # "2.1.4"

# 3. Upload a file
requests.post(f"{BASE}/sandboxes/{sandbox_id}/files", headers=HEADERS,
    files={"file": ("data.csv", open("data.csv", "rb"))},
    data={"path": "/workspace/data.csv"}
)

# 4. Process the file
result = requests.post(f"{BASE}/sandboxes/{sandbox_id}/processes", headers=HEADERS, json={
    "cmd": "python3",
    "args": ["-c", """
import pandas as pd
df = pd.read_csv('/workspace/data.csv')
print(f"Rows: {len(df)}, Columns: {list(df.columns)}")
print(df.describe().to_string())
"""]
}).json()

print(result["stdout"])

Templates

Choose a template based on your use case:

Template	Boot Time	Use Case
code_interpreter	~60ms	Data analysis, Python scripts
claude	~2s	Autonomous coding with Claude Code
codex	~2s	Code generation with OpenAI Codex
desktop	~5s	GUI automation, browser testing
base	~60ms	Custom environments

→ Full template details

Sandbox + Agent Integration

Attach a sandbox to an agent session for code execution capabilities:

# Create an agent with sandbox access
agent = requests.post(f"{BASE}/default/v1/agents", headers=HEADERS, json={
    "name": "Data Analyst",
    "model": {"name": "claude-sonnet-4"},
    "system": """You are a data analyst. When users ask questions about data,
write Python code to analyze it and execute it in your sandbox.
Always show your work and explain the results.""",
    "tools": [{"type": "mcp", "server": "code-exec"}]
}).json()

# Create a session with a sandbox environment
session = requests.post(f"{BASE}/default/v1/sessions", headers=HEADERS, json={
    "agent_id": agent["id"],
    "environment": {
        "template_id": "code_interpreter",
        "timeout": 600
    }
}).json()

# Now the agent can execute code in its sandbox
response = requests.post(
    f"{BASE}/default/v1/sessions/{session['id']}/events",
    headers=HEADERS,
    json={
        "type": "message",
        "role": "user",
        "content": "Analyze the sales data in /workspace/sales.csv and create a chart"
    }
).json()

Capabilities

Capability	API	Description
Execute commands	POST /sandboxes/:id/processes	Run any command with args
File read/write	GET/POST /sandboxes/:id/files	Upload, download, list files
Run to completion	POST /sandboxes/:id/processes	Synchronous; returns stdout/stderr + exit code
Pause/Resume	POST /sandboxes/:id/pause	Save state, resume later
Networking	Built-in	Outbound internet access

Limits

Resource	Limit
Boot time	~60ms (code_interpreter, base)
Max lifetime	1 hour
CPU	2 vCPU
Memory	4 GB
Disk	10 GB
Max concurrent	10 per account
File upload	100 MB

Next Steps

• Templates — Detailed specs for each template
• Sandbox Lifecycle — States, timeout, keepalive
• Sandbox API — Full API reference