Neura Market
knowlyr-sandbox logo

knowlyr-sandbox

Free

Code Agent 执行沙箱 — Docker 隔离、文件捕获、超时控制，为 LLM 代码代理提供安全的运行环境。

FreeFree tier
Type
Open Source
Visit Website GitHub (40 stars)

About knowlyr-sandbox

AgentSandbox is a reproducible Docker sandbox for code agent task execution and trajectory replay. It provides a standardized, isolated execution environment for LLM-based code agents, supporting task configuration via repository URL and commit hash, five standard tool interfaces (file read/write, shell execution, code search, git operations), and complete trajectory recording with replay capabilities. The sandbox can be used via CLI, Python API, or MCP server for integration with Claude and other AI assistants. It includes resource controls (CPU, memory, timeout) and snapshot/reset functionality.

Key Features

Docker container isolation for secure code execution
Five standard tool interfaces: file_read, file_write, shell, search, and git
Trajectory recording and replay for debugging and analysis
Snapshot and reset capability to restore sandbox to initial state
MCP server integration with Claude Desktop and Claude Code
Resource controls: CPU, memory, and timeout limits
CLI commands: create, exec, reset, replay, list
Supports Python pip installation with optional MCP and dev dependencies

Pros & Cons

Pros
  • Provides full Docker container isolation for security and reproducibility
  • Supports reproducible execution with fixed images and git commit snapshots
  • Offers five standard tool interfaces for code agent interaction
  • Enables trajectory replay for debugging and analysis
  • Integrates with Claude via MCP server
  • Includes CPU/memory/timeout resource controls
  • Open source and free to use
Cons
  • Requires Docker to be installed and running on the host
  • Currently only supports Python-based sandbox environment (python:3.11-slim shown in examples)
  • Limited to local execution; no cloud-hosted sandbox option
  • Complex setup for MCP server integration requires manual configuration
  • No built-in support for non-Docker execution environments

Best For

Running automated tests in isolated Docker containers for LLM agentsReplaying agent execution trajectories for debugging and verificationIntegrating with Claude via MCP for code agent tasksSecure execution of AI-generated code in a reproducible environmentEnabling reproducible development environments with fixed images and commit snapshots