AgentHubAgentHub

LLM Sandbox MCP Setup & Guide

MCP ServerMCP RegistryOfficial

io.github.vndee/llm-sandbox · v0.3.43

Securely run LLM-generated code in isolated containers across 7 languages and 3 container backends.

View on MCP Registry ↗Repository ↗

Copy the install config on this page first, then verify docs and permissions upstream.

Overview

Securely run LLM-generated code in isolated containers across 7 languages and 3 container backends. LLM Sandbox is a MCP Server listed from MCP Registry. Transports: stdio. This page includes an overview, setup tutorial, install commands, and use cases for Trae, Tongyi Lingma, Cursor, Claude Code, and VS Code.

Use cases

Code review & GitHubCursor essentials

AgentHub Verified AvailabilityTested & Ready

Automated pipeline validated install commands, protocol & client compatibility

Verified At2026-10-18
Install Snippets TestedCLI & JSON config syntax verified
Protocol Handshake ReadyComplies with JSON-RPC 2.0 specifications
Origin Registry ActiveSourced from official-mcp-registry
Verified ClientsClaude Code、Claude Desktop、Cursor 等
Security Tier: A+ 级 · 官方认证推荐 (A+)·Maintained by official/verified teams, audited for standard MCP protocol compliance.

Copy by platform

Choose your platform

  1. Open or create .cursor/mcp.json in your project root
  2. Click Copy config and paste; merge only this mcpServers entry if others exist
  3. Replace <placeholders> in env with real secrets (see Environment variables below)
  4. Save, then Cmd+Shift+P → Reload Window

Pre-fill Environment Variables (Optional)

Privacy Guarantee: Keys are replaced 100% locally in your browser memory and are NEVER uploaded or stored on any server.
Click to copy snippet
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Setup tutorial

  1. Open the LLM Sandbox page and confirm this MCP Server (source: MCP Registry).
  2. Copy the Cursor, Claude Code, or VS Code snippet.
  3. Merge it into mcpServers and replace env placeholders with real secrets.
  4. Reload the window, then call the MCP tools from your agent chat.

Install commands

Install commands and setup steps are in the HTML so search engines and no-JS browsers can read them without running client JavaScript.

Claude Code (local)

  1. Install Claude Code CLI
  2. Copy the command below, replace <placeholders> with real env values, then run in terminal
  3. See Environment variables below if listed
claude mcp add llm-sandbox -- uvx llm-sandbox

Cursor — .cursor/mcp.json (local)

  1. Open or create .cursor/mcp.json in your project root
  2. Click Copy config and paste; merge only this mcpServers entry if others exist
  3. Replace <placeholders> in env with real secrets (see Environment variables below)
  4. Save, then Cmd+Shift+P → Reload Window
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

VS Code — .vscode/mcp.json (local)

  1. Install GitHub Copilot in VS Code with MCP support
  2. Open or create .vscode/mcp.json in your project root
  3. Click Copy config and paste; merge only this mcpServers entry if others exist
  4. Replace <placeholders> in env with real secrets
  5. Save and Developer: Reload Window
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Claude Desktop — claude_desktop_config.json (local)

  1. Open Claude Desktop claude_desktop_config.json (see remote guide for paths)
  2. Click Copy config and merge under mcpServers
  3. Replace <placeholders> in env with real secrets
  4. Fully quit and restart Claude Desktop
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Trae — .trae/mcp.json (local)

  1. Trae → Settings → MCP, or edit .trae/mcp.json / global mcp.json
  2. Click Copy config and merge mcpServers
  3. Replace <placeholders> in env with real secrets
  4. Save and reload Trae
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Cherry Studio — MCP settings (local)

  1. Cherry Studio → Settings → MCP Servers → Add (STDIO)
  2. Or import JSON: click Copy config and merge mcpServers
  3. Replace <placeholders> in env; ensure Node.js / uv (npx, uvx) are installed
  4. Enable the server and check tools load
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Tongyi Lingma — MCP config (local)

  1. Lingma Settings → MCP → + → STDIO or config file
  2. Click Copy config and merge mcpServers
  3. Replace <placeholders> in env; need Node.js 18+ (npx) or uv (uvx)
  4. Confirm connected before using tools in agent chat
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Windsurf — mcp_config.json (local)

  1. Edit ~/.codeium/windsurf/mcp_config.json
  2. Click Copy config and merge mcpServers (stdio same as Cursor)
  3. Replace <placeholders> in env, save, refresh Cascade
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Cline — MCP Servers (local)

  1. Cline panel → Settings → MCP Servers
  2. Click Copy config and merge
  3. Replace <placeholders> in env, then save
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

WorkBuddy — .workbuddy/mcp.json (local)

  1. Edit ~/.workbuddy/mcp.json (user) or project .workbuddy/mcp.json
  2. Or Plugins → MCP Servers → Configure MCP and paste the JSON below
  3. Replace <placeholders> in env; on Windows prefer absolute paths for command/scripts
  4. Save, restart WorkBuddy, confirm connector status is green
{
  "mcpServers": {
    "llm-sandbox": {
      "command": "uvx",
      "args": [
        "llm-sandbox"
      ]
    }
  }
}

Troubleshooting & Common ErrorsFAQ

Common connection errors and verified fixes for LLM Sandbox

Getting 'connection closed' or exit code 1 in Cursor / Claude Code for LLM Sandbox?

Usually caused by missing runtime paths or IDE environment inheritance. Fix steps: 1. Verify Node.js 18+ (npx) or Python 3.10+ (uvx) is installed; 2. Run 'which npx' or 'which uvx' in terminal, and replace 'command' with absolute path; 3. Reload or restart the client window after modifying config.

Fix Snippet
# Check binary path in terminal:
which npx
node -v
Getting 'spawn npx ENOENT' or 'command not found'?

The editor background process does not inherit your full terminal PATH. Solution: Globally install the package, or set the absolute binary path (e.g. C:\Program Files\nodejs\npx.cmd on Windows, or /usr/local/bin/npx on macOS).

Missing required environment variable or authentication failure?

LLM Sandbox requires environment variables (BACKEND、DOCKER_HOST、KUBECONFIG、NAMESPACE、COMMIT_CONTAINER、KEEP_TEMPLATE、SANDBOX_NETWORK_MODE、SANDBOX_READ_ONLY、SANDBOX_CAP_DROP、SANDBOX_SECURITY_OPT、SANDBOX_MEMORY、SANDBOX_CPUS). Ensure you have added valid keys under the 'env' object in your config file without trailing whitespace.

Fix Snippet
// .cursor/mcp.json 或 claude_desktop_config.json
{
  "env": {
    "API_KEY": "your_actual_key_here"
  }
}
Still not connecting? Read the MCP troubleshooting guide →

Tool Mock Playground

Sandbox

LLM Sandbox Tool Interface Simulation · Preview tool schema & outputs without local runtime

Sandbox Ready
fn: llm_sandboxExecute the core tool interface of LLM Sandbox
Request ArgumentsJSON Schema
{
  "target": "LLM Sandbox",
  "action": "execute",
  "options": {
    "mode": "standard",
    "timeoutMs": 5000
  }
}
💡Parameters generated dynamically by Agent runtime
Agent Tool Output

Click 'Run Mock' above

to preview the raw response returned to the LLM

Env: AgentHub Virtual SandboxJSON-RPC 2.0

Decision Guide: Why & When to Use

Assess suitability before installing to save trial-and-error time

Best Suited For
  • PR review
  • Changelog generation
  • Cross-repo issue search
When NOT to Use
  • Replacing human security audit
  • Unauthorized repo access
Recommended Workflow Pairing:View Scenario →

LLM Sandbox + Scenario Prompt → Complete Agent Automation

MCP hands-on: install to visible results

Follow the full lab (expected UI/output + contrast checks). After installing this item, verify with the tutorial prompts.

Open tutorial →

Related resources

Often paired with

中文公文写作

v2.0.41

SkillClawHub10.2k

io.clawhub.gongyu0918-debug/chinese-official-writing

用于中文公文、事务性材料和新闻稿件的起草、改写、压缩、润色、审校、文种核对、去口语化、降 AI 味及 Word 格式处理，适用于机关、企事业单位、学校和新闻机构。涵盖申请、请示、报告、通知、通告、意见、决定、决议、议案、公报、命令、函、复函、批复、说明、方案、纪要、公告、公示、通报、制度、规定、办法、细则、操作规程、工作要点、总结、调研、讲话、致辞、主持词、述职、可研、审查材料、技术需求、新闻消息、编者按、新闻评论，以及采购、整改、反馈和 AI 算力等场景。

source
View on ClawHub

Planning with files

v3.25.0

SkillClawHub45.4k

io.clawhub.othmanadi/planning-with-files

Persistent file-based planning for multi-step AI-agent work. Keeps task_plan.md, findings.md, and progress.md on disk; lifecycle hooks inject selected project planning context. Automatic recovery reads project planning files only. Explicit session-catchup.py --metadata reads same-project local agent session records and emits aggregate counts only; --replay may emit bounded nonce-framed excerpts. Optional gated mode can request continuation only when the host supports it and never runs commands declared in Markdown. The skill has no network upload path. Use for research or work needing 5+ tool calls.

source
View on ClawHub

Keep exploring AgentHub

Most people compare similar tools or check scenario guides before installing—start here.

Browse more like thisSearch similarPick by scenarioPrompt playbooksMCP hands-onReviews & comparisons

Listing badge: put AgentHub on your site

Copy either snippet into your project homepage, docs, or GitHub README. The badge is a hotlinked SVG — nothing to host — and it links back to this page so visitors can find the install steps.

PreviewListed on AgentHub: LLM Sandbox
HTML
<a href="https://myagenthub.cn/p/io.github.vndee/llm-sandbox" title="Listed on AgentHub: LLM Sandbox" target="_blank" rel="noopener">
  <img src="https://myagenthub.cn/badge/io.github.vndee/llm-sandbox?lang=en" alt="Listed on AgentHub: LLM Sandbox" height="20" style="border:0"/>
</a>
Markdown (GitHub README)
[![Listed on AgentHub: LLM Sandbox](https://myagenthub.cn/badge/io.github.vndee/llm-sandbox?lang=en)](https://myagenthub.cn/p/io.github.vndee/llm-sandbox)

Badges are generated on the fly from /badge/<package-id>, so name and listing changes propagate automatically. Keep the link target unchanged — it is what counts as the referral.

Unified Manifest

{
  "id": "io.github.vndee/llm-sandbox",
  "type": "mcp-server",
  "version": "0.3.43",
  "displayName": "LLM Sandbox",
  "description": "Securely run LLM-generated code in isolated containers across 7 languages and 3 container backends.",
  "repository": {
    "url": "https://github.com/vndee/llm-sandbox",
    "source": "github"
  },
  "homepage": "https://vndee.github.io/llm-sandbox/",
  "distribution": {
    "packages": [
      {
        "registryType": "pypi",
        "identifier": "llm-sandbox",
        "version": "0.3.43",
        "runtimeHint": "uvx",
        "transport": "stdio",
        "environmentVariables": [
          {
            "name": "BACKEND",
            "description": "Container backend to use. Must match the installed extra: mcp-docker, mcp-podman, or mcp-k8s."
          },
          {
            "name": "DOCKER_HOST",
            "description": "Docker or Podman socket URL, e.g. unix:///var/run/docker.sock"
          },
          {
            "name": "KUBECONFIG",
            "description": "Path to kubeconfig file when BACKEND=kubernetes."
          },
          {
            "name": "NAMESPACE",
            "description": "Kubernetes namespace used for sandbox pods when BACKEND=kubernetes."
          },
          {
            "name": "COMMIT_CONTAINER",
            "description": "Commit the container after a run so installed libraries persist between sessions."
          },
          {
            "name": "KEEP_TEMPLATE",
            "description": "Keep the base image after the session ends to avoid re-pulling it on the next run."
          },
          {
            "name": "SANDBOX_NETWORK_MODE",
            "description": "Network mode for the sandbox container. Set to 'none' for hardened isolation. Docker and Podman backends only."
          },
          {
            "name": "SANDBOX_READ_ONLY",
            "description": "Mount the sandbox root filesystem read-only. Recommended: true. Docker and Podman backends only."
          },
          {
            "name": "SANDBOX_CAP_DROP",
            "description": "Comma-separated Linux capabilities to drop. Recommended: ALL. Docker and Podman backends only."
          },
          {
            "name": "SANDBOX_SECURITY_OPT",
            "description": "Comma-separated container security options, e.g. no-new-privileges. Docker and Podman backends only."
          },
          {
            "name": "SANDBOX_MEMORY",
            "description": "Memory limit for the sandbox container, e.g. 4g. Docker and Podman backends only."
          },
          {
            "name": "SANDBOX_CPUS",
            "description": "Fractional CPU allocation for the sandbox container, e.g. 1.5. Docker and Podman backends only."
          }
        ]
      }
    ],
    "remotes": []
  },
  "dependencies": [],
  "installTargets": [
    "claude-code",
    "claude-desktop",
    "cursor",
    "vscode",
    "trae",
    "cherry-studio",
    "lingma",
    "windsurf",
    "cline",
    "workbuddy"
  ],
  "keywords": [],
  "provenance": {
    "origin": "official-mcp-registry",
    "originalId": "io.github.vndee/llm-sandbox",
    "originalUrl": "https://registry.modelcontextprotocol.io/v0.1/servers/io.github.vndee%2Fllm-sandbox/versions/latest",
    "isOfficial": true,
    "status": "active"
  }
}
LLM Sandbox MCP Setup & Guide - AgentHub