calling-mcp-tools-via-subprocess

name	calling-mcp-tools-via-subprocess
description	Bypasses a flaky MCP broker by spawning the server directly.
tier	optional
category	integration
created_by	human
platforms	["windows","macos","linux"]
tags	["mcp","subprocess","reliability"]
author	Andreas Wasita (@andreaswasita)

Calling MCP Tools via Subprocess Skill

Bypasses a flaky MCP broker by spawning the server binary directly and speaking JSON-RPC over stdio, using the dojo helpers mcp/scripts/mcp-subprocess.js and mcp-subprocess.ps1. Does NOT replace using-mcp for normal operation — broker is the fast path; subprocess is the reliable fallback.

When to Use

• Broker shows the server as connected but tools/call hangs or times out.
• Building a script (CI job, scheduled refresh, batch backfill) that needs MCP data without an IDE.
• Deterministic agent workflows that must succeed regardless of client flakiness.
• Diagnosing whether a problem is in the broker or in the server itself.
• User says "broker stuck", "tools not loading", "tools_changed notice".

Prerequisites

• The dojo helpers mcp/scripts/mcp-subprocess.js and mcp/scripts/mcp-subprocess.ps1.
• The MCP server binary resolvable (npm install or known path).
• Node 18+ on the path.
• Required auth context set in env BEFORE spawning (e.g. az login, gh auth).
• The powershell Copilot tool to run the helper.

How to Run

1. Resolve the server binary path.
2. Call the helper with binary, tool name, and JSON args.
3. Receive parsed `result.content` text.
4. Wrap recurring calls in a per-server module to keep skill code clean.
5. Cache results in-session (or to `.cache/mcp/`) since startup is 200–800ms.

Quick Reference

Helper	Use from
node mcp/scripts/mcp-subprocess.js call <bin> <tool> <jsonArgs>	bash / cmd / CI
Invoke-McpTool -Bin <bin> -Tool <name> -Params @{...}	PowerShell

Binary location	Source
~/.copilot/installed-plugins/_direct/<vendor>--<name>/dist/index.js	Copilot CLI
~/.npm/_npx/<hash>/node_modules/<pkg>/dist/index.js	npx cache
Global install: npm i -g <pkg> then which <pkg>	Most reliable

Auth source	Set before spawn
Azure / MSX	az login
GitHub	gh auth login OR $env:GITHUB_PERSONAL_ACCESS_TOKEN
Kubernetes	Current kubeconfig context

Procedure

Step 1: Resolve the Binary

Pick the most reliable location from the table. For production scripts, npm i -g <pkg> once and pin the version — do not depend on the npx cache.

Step 2: Call the Helper

bash:

node mcp/scripts/mcp-subprocess.js call "$BIN" get_my_deals '{"limit":10}'

PowerShell:

. mcp/scripts/mcp-subprocess.ps1
Invoke-McpTool -Bin "C:\path\to\server\index.js" -Tool "get_my_deals" -Params @{ limit = 10 }

Both helpers do exactly:

1. Spawn node <bin> with piped stdio.
2. Send initialize (id=1).
3. Send tools/call (id=2).
4. Parse the response line containing "id":2.
5. Return result.content text.
6. Time out after 60s and kill the child.

Step 3: Wrap Per Server

Do not sprinkle subprocess calls through skills. Wrap each server in one module:

// scripts/github-mcp.js
const { callTool } = require("../mcp/scripts/mcp-subprocess.js");
const BIN = require("path").join(process.env.HOME, ".npm", "_npx", "...", "server-github", "dist", "index.js");
exports.searchRepos = (q) => callTool(BIN, "search_repositories", { query: q });

Step 4: Cache

Subprocess startup costs 200–800ms per call. For multi-call flows:

• Batch related calls into one subprocess run if the server supports it.
• Cache in-session keyed by (tool, JSON.stringify(args)).
• Persist to .cache/mcp/<server>.json for long-running scripts.

Step 5: Timeouts and Errors

• Use the 60s default; override only for known-slow tools.
• Never block forever.
• On failure, log captured stderr — the truth is there.

Step 6: Fallback Wrapper

async function callWithFallback(toolName, args) {
  try {
    return await nativeMcpCall(toolName, args);
  } catch (err) {
    console.error(`Broker failed (${err.message}); falling back to subprocess.`);
    return await callTool(BIN, toolName, args);
  }
}

Test the subprocess path in CI even if your IDE uses the broker.

Pitfalls

• DO NOT re-implement JSON-RPC framing inline. Use the helper.
• DO NOT rely on npx -y per invocation in production scripts. Install once.
• DO NOT swallow errors. Always surface stderr.
• DO NOT keep a single child process open for the whole session pushing many requests through it — response correlation races will bite.
• DO NOT pass credentials via tool arguments. Set env before spawn.

Verification

• Helper invocation returns the expected tool output (text content) for at least one call.
• Server binary path is resolved deterministically (not from a transient cache).
• Timeout is set (default 60s, or explicit override).
• Wrapper module exists per server — no raw callTool scattered across skills.
• CI exercises the subprocess path independently of the broker.
• No secrets passed through tool arguments (grep the call sites).