Skip to content

vinmay/reachscan

v0.1.0Apache-2.0

Vet an MCP server before you install it: see what each tool can execute, read, write, and send, using reachscan's static reachability analysis.

GitHubOpen Website

reachscan agent plugin (Claude Code and Codex)

One skill, vet-mcp-server, packaged for both Claude Code and Codex. Before you add an MCP server, ask your agent to vet it. The agent runs the reachscan CLI on the server and tells you what each tool can execute, read, write, and send, then gives an Install / Review first / Avoid verdict based only on the scan findings.

> vet this MCP server before I add it: https://github.com/org/some-mcp-server

The skill calls the reachscan CLI directly. There's no MCP server and no network service. If reachscan isn't installed, the skill tells you to run pipx install reachscan and stops. It never installs anything on its own.

Requirements

  • reachscan on your PATH: pipx install reachscan
  • Python 3.11 or later (for reachscan)
  • An agent that can run shell commands, such as Claude Code or Codex. In chat apps without a terminal, the skill explains that it needs one instead of guessing.

Install in Claude Code

From inside Claude Code:

/plugin marketplace add vinmay/reachscan
/plugin install reachscan@reachscan

Or from a local checkout, for testing:

claude --plugin-dir ./integrations/agent-plugins

Install in Codex

Add this repo as a marketplace:

codex plugin marketplace add vinmay/reachscan

Then open the Plugins Directory, choose the reachscan marketplace, and install the plugin. The repo marketplace file is .agents/plugins/marketplace.json at the repo root.

What's in here

integrations/agent-plugins/
├── plugin.json                  # portable Agent Plugins manifest, with OpenAI metadata under extensions.com.openai
├── .claude-plugin/plugin.json   # Claude Code manifest
├── .codex-plugin/plugin.json    # Codex compatibility manifest, for older Codex clients
├── assets/icon.svg              # listing icon
└── skills/vet-mcp-server/SKILL.md

Marketplace files at the repo root:

  • .claude-plugin/marketplace.json for Claude Code
  • .agents/plugins/marketplace.json for Codex

Supported targets

GitHub URLs, pypi:name[==version], local paths, and mcp+https:// endpoints. npm packages aren't supported yet.

Limits

reachscan analyzes Python tool bodies. TypeScript/JavaScript tool handlers are detected, but their bodies aren't analyzed yet, and the skill says so when it matters. The verdict reflects code patterns, not proven runtime behavior.

Privacy

reachscan runs locally and sends no data anywhere. See PRIVACY.md.

Building the OpenAI submission ZIP

From the repository root:

python scripts/build_openai_plugin_zip.py

This writes dist/reachscan-plugin-<version>.zip (the plugin folder plus the license, without the Claude Code manifest) for upload at platform.openai.com/plugins.