safe-agent-mcp

A secure multi-agent MCP gateway based on DeepSeek Harness: it wraps DSH's agent execution capabilities into standard MCP tools, so any MCP client (Claude Desktop, Cursor, etc.) can call them.

What this project solves

• DSH is a multi-agent framework, but its capabilities can only be used in its own interface;

• MCP is a standard protocol that any client can connect to, but it has no agent capabilities of its own;

• safe-agent-mcp is the bridge between the two: an MCP server that provides standard tools externally and drives DSH's agents internally (including sub-agents and the sandbox).

Related MCP server: py-sandbox

Features

Quick Start

npm install
npm run smoke     # 自动化冒烟测试（协议握手 + 工具调用）
npm run inspect   # 打开 MCP Inspector 可视化调试

Currently implemented tools:

• run_command: executes a shell command in the sandbox and returns stdout / stderr / exit code (the GBK encoding issue of cmd is handled on Windows).

Directory Structure

src/index.ts        # MCP 服务器入口
tests/smoke.mjs     # 冒烟测试
docs/DEVELOPMENT.md # 开发文档（架构、里程碑、开发日志）

Milestones

• v1.0: MCP server + run_command tool (smoke tests all green)

• v1.1: run_agent_task multi-agent integration (headless bridge)

• v1.2: Security shell (dangerous command interception + secret redaction + audit logs)

• v1.3: Web dashboard (bonus)

See docs/DEVELOPMENT.md for details.

License

MIT