Claude Code

We Built an MCP Server That Turns Claude Into Your Claude Code Config Guide

The @vibetokens/mcp server installs in 60 seconds and puts Claude Code configuration guidance — audit, hidden features, coordinator mode — directly inside every session.

MurphMarch 31, 20266 min read

The problem with most configuration guides is that they're outside your workflow.

You're in the middle of a session, something feels off, and you'd have to open a browser, navigate to a tool, paste your config, read the output, then figure out what to do next — while your session is still running.

We wanted to solve this the right way: put the guidance inside the session itself.


What the MCP server does

@vibetokens/mcp is a Model Context Protocol server with four tools. Once installed, Claude can call any of them automatically when you ask about your setup, configuration, or Claude Code features.

audit_settings — Pass your ~/.claude/settings.json. Claude returns a 0–6 power score, a list of what's passing, and specific fixes for every gap. Same engine as the browser audit tool.

get_hidden_features — The complete reference: every confirmed CLI flag, environment variable, hidden slash command, and coordinator mode pattern from Claude Code v2.x source. Not in official docs. All in one tool call.

get_claude_code_tips — Actionable patterns for CLAUDE.md architecture, coordinator mode, hooks, and the --effort flag. Ask about a specific topic or get the full set.

get_operator — Full OPERATOR details: what's in the kit, what each component does, pricing, install steps.


Why this distribution pattern matters

Most developer tools think of the product and the documentation as separate things. The product lives at some URL. The docs live at another URL. You navigate between them.

MCP servers collapse that gap. The product becomes a tool that lives inside the AI layer. You don't leave your session to get guidance — you ask Claude, and Claude asks the tool.

This is the model. Not "here's a guide, read it." But "the AI is already in the loop. Let it surface the right information at the right moment."

We built the browser audit tool first because it's fast to ship and useful on its own. But the MCP server is the right architecture for the problem. The AI becomes the interface.


Install

Add to ~/.claude/claude_desktop_config.json (or your Cursor MCP config):

{
  "mcpServers": {
    "vibetokens": {
      "command": "npx",
      "args": ["-y", "@vibetokens/mcp"]
    }
  }
}

Restart Claude. Then just ask:

Check my Claude Code settings — here's my settings.json: [paste]
What Claude Code features am I missing?
How do I activate coordinator mode?

Claude picks up the context and calls the right tool automatically.


What it finds

The audit checks six configurations that most setups are missing:

  • AUTOCOMPACT_PCT_OVERRIDE — The default context compaction threshold is ~70%. Every long session starts trimming memory before it has to. Setting this to 88 buys you significantly more working context before anything gets cut.

  • MAX_THINKING_TOKENS — The internal reasoning budget per response. Default is lower than optimal. Set this to 16000 and the difference is noticeable on anything architecturally complex.

  • IDLE_THRESHOLD_MINUTES — How long before Claude considers a session stale. Default is too low for real work sessions. 60 minutes covers a normal working hour.

  • skipDangerousModePermissionPrompt — Without this set with an explicit allow list, every tool call interrupts automation flows. If you're running any kind of autonomous pipeline, this is a hard requirement.

  • permissions.allow list — What runs automatically without a manual approval. Missing this means every Bash, Read, Write, Edit, Glob, and Grep call is gated.

  • hooks — Stop, PreToolUse, PostToolUse event hooks. The automation layer most people skip entirely. A Stop hook that plays a sound when Claude finishes means you never have to watch the terminal again.

Zero out of six is common. Most developers install Claude Code, run it, and never configure it. The defaults are functional but not optimized.


The scores

The audit uses a 0–6 scoring model:

  • 0–1: Unconfigured
  • 2–3: Basic
  • 4–5: Functional
  • 6: Power User

Most setups come in at 1 or 2. The fixes are usually 10 minutes of copy-paste work. OPERATOR is the drop-in kit that gets you to Power User in 15 minutes flat — correct settings pre-configured, master CLAUDE.md template with coordinator mode wired, PowerShell aliases, automation scripts, and the full hidden features reference.


Why the MCP pattern compounds over time

We're going to add more tools to this server. More checks, more guidance surfaces, more ways for Claude to bring relevant information into the session at the right moment.

Every new tool compounds the value of having installed the server. Once it's in your config, you get every update automatically — next time you run npx -y @vibetokens/mcp, you get the latest version.

The browser audit tool is a standalone snapshot. The MCP server is a continuously improving layer inside Claude.

Free. MIT. One JSON block to install.

vibetokens.io/mcp — install docs, tool reference, example prompts → npm: @vibetokens/mcp

Want to see how your business stacks up?

Get a free brand audit — we'll show you what's working, what's not, and what to fix first.

Free Brand Audit →

Frequently Asked

What is the @vibetokens/mcp server?

It's a free Model Context Protocol server that gives Claude access to four tools: audit_settings (audits your settings.json), get_hidden_features (complete reference of undocumented flags and env vars), get_operator (full OPERATOR kit details), and get_claude_code_tips (actionable patterns for CLAUDE.md, coordinator mode, hooks, and the effort flag). Install it once — Claude can use these tools automatically when you ask about your Claude Code setup.

How do I install the @vibetokens/mcp server?

Add one JSON block to ~/.claude/claude_desktop_config.json or your Cursor MCP config: { "mcpServers": { "vibetokens": { "command": "npx", "args": ["-y", "@vibetokens/mcp"] } } }. Restart Claude. Done. No npm install required — npx handles it on first run.

What does the MCP server audit check?

Six configurations: AUTOCOMPACT_PCT_OVERRIDE (context retention), MAX_THINKING_TOKENS (reasoning depth), IDLE_THRESHOLD_MINUTES (session longevity), skipDangerousModePermissionPrompt (automation bypass), permissions.allow list, and hooks setup. Returns a 0–6 power score with a specific fix for each gap.

What's the difference between the MCP server and the browser audit tool?

Same audit engine, different surface. The browser tool at vibetokens.io/tools/claude-code-audit is a standalone web app — you visit it, paste your settings.json, get a report. The MCP server lives inside Claude — you stay in your session and just ask Claude to check your setup. Claude calls the tool automatically.

Jason Murphy

Written by

Murph

Jason Matthew Murphy. Twenty years building digital systems for businesses. Former CardinalCommerce (acquired by Visa). Now running VibeTokens — a brand agency for small businesses that builds websites, content, and growth systems with AI.

Your brand is your first impression.

Find out if it's costing you customers.

Free brand audit. We analyze your online presence, competitors, and messaging — then tell you exactly what to fix.

Get Your Free Brand Audit →