vital-agentbox 0.1.5

Secure sandboxed code execution and agent toolbox

vital-agentbox

Secure sandboxed code execution for AI agents. Runs Python and shell commands inside a Chromium + Pyodide (WASM) sandbox with two independent security boundaries — no host filesystem or network access from agent code.

Features

• Dual-layer isolation — Chromium renderer sandbox + WASM linear memory
• Virtual shell — tree-sitter-bash parser with 30+ builtins on in-memory FS
• Python execution — Pyodide (CPython 3.11 compiled to WASM)
• Git operations — isomorphic-git on Emscripten MemFS with S3/MinIO storage
• AI-friendly editing — edit builtin with fuzzy + AST-aware matching
• LangChain & Deep Agents — toolkit, tools, and sandbox backend integrations
• Scalable — orchestrator + worker architecture with Redis routing

Box types

Type	Description
MemBox	Ephemeral in-memory sandbox (default)
GitBox	MemBox + isomorphic-git + pluggable storage (S3/MinIO/local)
FileSystemBox	Local dev only, backed by host directory

Install

# Lightweight client (for LangGraph / Deep Agent apps)
pip install vital-agentbox[client]

# Sandbox worker (runs Chromium + Pyodide)
pip install vital-agentbox[worker]
playwright install chromium

# Orchestrator (routes requests to workers, no Chromium)
pip install vital-agentbox[orchestrator]

# LangChain integration
pip install vital-agentbox[langchain]

Quick start

from agentbox.client import AgentBoxClient

client = AgentBoxClient("http://localhost:8090")

# Create a sandbox
sandbox = client.create_sandbox_sync(box_type="mem")

# Run Python
result = sandbox.execute_sync("print(2 + 2)")
print(result.stdout)  # "4\n"

# Run shell commands
result = sandbox.execute_sync('echo "hello" > /file.txt && cat /file.txt', language="shell")
print(result.stdout)  # "hello\n"

# AI-friendly file editing
result = sandbox.execute_sync(
    "edit /file.txt --old 'hello' --new 'world'",
    language="shell",
)

# Cleanup
sandbox.destroy_sync()

LangChain integration

from agentbox.langchain import AgentBoxToolkit

toolkit = AgentBoxToolkit(base_url="http://localhost:8090")
tools = toolkit.get_tools()
# → [CodeExecutionTool, ShellExecutionTool, FileWriteTool, FileReadTool]

Docker

# Full stack (orchestrator + 2 workers + MinIO)
docker compose up

# Single worker
docker run -p 8090:8000 --shm-size=2g agentbox-worker

Documentation

Full documentation is in the docs/ directory:

• Getting started
• Sandbox overview
• Shell builtins reference
• Client SDK reference
• REST API reference
• Deployment guide

System requirements

• Python ≥ 3.11
• Chromium (via playwright install chromium) — worker only
• Redis — orchestrator only
• For PDF generation: pandoc + LaTeX

License

Apache 2.0