mcp-netassist

Network and proxy diagnostics your agent can run itself — in any MCP client: Claude Code, Cursor, Claude Desktop or DSH.

When you need this

• Your MCP client cannot reach GitHub or npm and you want the agent to find out why
• You want the same diagnostics in every editor instead of one plugin per host
• You are tired of the agent guessing at network problems

The score badge says it plainly: this server is listed on Glama, claimed by its maintainer, with its tool definitions rated A. Requires Windows — see Platform below.

Works with any MCP client — Claude Code, Claude Desktop, Cursor, Reasonix, CodeWhale, DeepSeek Harness. Point your agent at it and ask "is GitHub reachable?", "why is my proxy not working?", "what should I change?" — instead of guessing.

Built for the China-network reality: flaky GitHub, proxies that are half-configured, hosts files that fight the proxy, and TUN mode that silently overrides the system proxy.

Windows-only for now: the checks call PowerShell. The protocol layer is portable; a POSIX backend is the obvious next step.

Feedback

Installs are visible in the download counts; whether this helps is not. If you use it, one line about the actual task is worth more than a star — particularly if it did not work, which is the case that never gets reported:

• what you were trying to do
• whether it did it
• what was missing or wrong

Tell me in an issue — or if something is already broken, just open a normal bug report.

Platform

Windows. The checks read the system proxy from the Windows registry and shell out to powershell.exe; on Linux and macOS the tools start but their checks cannot run, and they report PowerShell failed: spawn powershell.exe ENOENT rather than pretending to have checked something.

The packaging is portable (any MCP client can connect, the server speaks plain stdio), but the diagnostics are Windows-specific today. A POSIX implementation would read the proxy from the environment and use ss/lsof for port probing; that is not written yet, so the README says Windows instead of implying otherwise.

Timing

Every check spawns a powershell.exe, and a cold one costs roughly 20 seconds to start. So on a freshly booted Windows machine:

• the individual tools answer within about half a minute
• net_doctor can exceed a minute, because it runs four checks in parallel plus a port probe

Most MCP clients default to a 60-second request timeout, so net_doctor may time out on a cold machine even though it is working. If that happens, call the individual checks (net_github_status, net_proxy_status, net_hosts_check) instead, or raise your client's request timeout. Once PowerShell has been used a few times the cost drops sharply.

Tools

net_doctor is the point of this server. Other tools tell you what is wrong; it tells you what to do about it:

✔ System proxy: 127.0.0.1:10808
✔ Proxy port 10808 responding
✔ GitHub reachable (HTTP 200, 312 ms)
⚠ TUN-style adapter detected: clash
   Under TUN mode the system proxy setting is usually ignored — the two can fight each other.
✔ hosts file clean (no GitHub entries)

Suggested fix:
- Under TUN mode, clear the Windows system proxy (or exclude github.com) so traffic is not double-handled.

Setup

Claude Desktop / Cursor / any JSON-configured client:

{
  "mcpServers": {
    "netassist": {
      "command": "npx",
      "args": ["-y", "github:Edge-Echo/mcp-netassist"]
    }
  }
}

The built lib/ ships in the repository, so the GitHub form needs no build step.

From npm:

{
  "mcpServers": {
    "netassist": {
      "command": "npx",
      "args": ["-y", "mcp-netassist"]
    }
  }
}

Claude Code:

claude mcp add netassist -- npx -y github:Edge-Echo/mcp-netassist

DeepSeek Harness (same diagnostics as a native plugin, plus net_doctor as an agent tool):

dsh plugin --profile web add dsh-netassist

From a checkout:

{
  "mcpServers": {
    "netassist": { "command": "node", "args": ["/path/to/mcp-netassist/lib/server.js"] }
  }
}

In a container:

docker build -t mcp-netassist .
docker run -i --rm mcp-netassist

The image is also what directory listings use for introspection checks. Inside a Linux container the server starts and answers initialize / tools/list normally; the tools themselves need Windows PowerShell, and say so when it is missing.

Design notes

• Read-only. No tool writes config, changes the proxy, or edits the hosts file. It reports and suggests; you decide.
• Injection-safe. Every user input crosses the PowerShell boundary as Base64, never as interpolated text.
• No hidden state. Each call runs its own checks; nothing is cached between calls, so results are always current.
• Two dependencies (@modelcontextprotocol/sdk, zod), no native modules.

Related

Part of the dsh-toolkit family — the same diagnostics also ship as a DeepSeek Harness plugin (dsh-netassist), which adds net_doctor as an agent tool.

License

MIT