Secure Agent Skills Execution

Acontext Sandbox provides secure, isolated environments for code and file management.

Open Source Alternative to Claude Skills API

Simple & Open Source

Easy to integrate, community-driven

Model Agnostic

Works with any LLM

Composable

Mix tools, skills, sandboxes

How Sandbox Works with LLM

Give your AI agent secure code execution capabilities through an agentic tool loop

setup.py
# Create isolated sandbox & persistent disk
sandbox = client.sandboxes.create()
disk = client.disks.create()
# Upload agent skill (pptx generator, etc.)
skill = client.skills.create(file=zip_file)
📦Sandbox Ready
sandbox_id:"sbx_7k2m9x"
env:Python 3, bash, ripgrep...
💾Disk Created
disk_id:"dsk_3f8n2p"
storage:"persistent"
🧩Skill Uploaded
skill_id:"skl_pptx_v1"
mount_path:"/skills/pptx/"
context.py
# Build context with tools & mounted skills
ctx = SANDBOX_TOOLS.format_context(
  client, sandbox_id, disk_id,
  mount_skills=[skill.id]
)
# Get OpenAI-compatible tool schema
tools = SANDBOX_TOOLS.to_openai_tool_schema()
📄System Prompt Includes
🔧Available tools: bash, editor, export
📁Mounted skills at /skills/
📋Skill instructions & examples
🔧Tools → LLM
bash_execution_sandbox
Run shell commands
text_editor_sandbox
View/create/edit files
export_file_sandbox
Export with public URL
👤User Request
"Create a sales report presentation"
🧠
LLM
Decides tool call
Tool Call
bash_execution_sandbox
{cmd: "python3 /skills/pptx/main.py"}
📦
Sandbox
Executes securely
✅Tool Result
stdout: "Presentation created at /output/report.pptx"
↻Loop until done
export.py
# LLM calls export tool to share results
if tool_name == "export_file_sandbox":
  result = SANDBOX_TOOLS.execute_tool(
    ctx, tool_name, {"sandbox_path": path}
  )
🔗Public URL Generated
{
  "public_url": "https://cdn.acontext.io/f/report.pptx",
  "expires_in": "24h",
  "file_size": "2.4 MB"
}
🤖Assistant
Done! Here's your presentation: cdn.acontext.io/f/report.pptx

Acontext Skills Execution vs Claude Skills API

Compare the complexity of traditional approaches with Acontext's unified SDK

Code Execution & Skills

Execute code in sandboxes and use agent skills without complex beta APIs

ComplexClaude API Skills

SimpleAcontext Sandbox

# Acontext SDK - Simple & unified
from acontext import AcontextClient
from acontext.agent.sandbox import SANDBOX_TOOLS
from openai import OpenAI

client = AcontextClient()
openai = OpenAI()

# Create sandbox with skill mounted
sandbox = client.sandboxes.create()
disk = client.disks.create()

ctx = SANDBOX_TOOLS.format_context(
    client,
    sandbox_id=sandbox.sandbox_id,
    disk_id=disk.id,
    mount_skills=["excel-skill-uuid"]  # Mount any skill
)

# Use with any LLM - OpenAI, Anthropic, etc.
response = openai.chat.completions.create(
    model="gpt-4.1",
    messages=[
        {"role": "system", "content": ctx.get_context_prompt()},
        {"role": "user", "content": "Create an Excel budget spreadsheet"}
    ],
    tools=SANDBOX_TOOLS.to_openai_tool_schema()
)

# Execute tool and get result with download URL
for tc in response.choices[0].message.tool_calls:
    result = SANDBOX_TOOLS.execute_tool(
        ctx, 
        tc.function.name, 
        json.loads(tc.function.arguments)
    )
    # Result includes public_url for downloads

# Claude API with Skills - Complex setup
import anthropic

client = anthropic.Anthropic()

# Step 1: Create message with Skills in container
response = client.beta.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=4096,
    betas=["code-execution-2025-08-25", "skills-2025-10-02"],
    container={
        "skills": [
            {"type": "anthropic", "skill_id": "xlsx", "version": "latest"}
        ]
    },
    messages=[{
        "role": "user",
        "content": "Create an Excel file with a budget spreadsheet"
    }],
    tools=[{"type": "code_execution_20250825", "name": "code_execution"}]
)

# Step 2: Extract file IDs from nested response structure
def extract_file_ids(response):
    file_ids = []
    for item in response.content:
        if item.type == 'bash_code_execution_tool_result':
            content_item = item.content
            if content_item.type == 'bash_code_execution_result':
                for file in content_item.content:
                    if hasattr(file, 'file_id'):
                        file_ids.append(file.file_id)
    return file_ids

# Step 3: Download files using separate Files API
for file_id in extract_file_ids(response):
    file_metadata = client.beta.files.retrieve_metadata(
        file_id=file_id,
        betas=["files-api-2025-04-14"]
    )
    file_content = client.beta.files.download(
        file_id=file_id,
        betas=["files-api-2025-04-14"]
    )
    file_content.write_to_file(file_metadata.filename)

# Acontext SDK - Simple & unified
from acontext import AcontextClient
from acontext.agent.sandbox import SANDBOX_TOOLS
from openai import OpenAI

client = AcontextClient()
openai = OpenAI()

# Create sandbox with skill mounted
sandbox = client.sandboxes.create()
disk = client.disks.create()

ctx = SANDBOX_TOOLS.format_context(
    client,
    sandbox_id=sandbox.sandbox_id,
    disk_id=disk.id,
    mount_skills=["excel-skill-uuid"]  # Mount any skill
)

# Use with any LLM - OpenAI, Anthropic, etc.
response = openai.chat.completions.create(
    model="gpt-4.1",
    messages=[
        {"role": "system", "content": ctx.get_context_prompt()},
        {"role": "user", "content": "Create an Excel budget spreadsheet"}
    ],
    tools=SANDBOX_TOOLS.to_openai_tool_schema()
)

# Execute tool and get result with download URL
for tc in response.choices[0].message.tool_calls:
    result = SANDBOX_TOOLS.execute_tool(
        ctx, 
        tc.function.name, 
        json.loads(tc.function.arguments)
    )
    # Result includes public_url for downloads