Skip to content

Claude Code Ultimate Guide MCP server

Last updated:

Technical reference for claude-code-ultimate-guide-mcp, the local stdio server that exposes this guide to MCP-compatible coding clients.

Status checked on 2026-09-01:

  • public npm package: 1.3.1
  • public runtime handshake: 1.3.1
  • public runtime surface: 17 tools, 6 resources, 1 prompt
  • repository package version: 1.3.1
  • official MCP Registry: io.github.FlorianBruniaux/claude-code-guide, active at version 1.3.1

The public-package facts above were measured by starting claude-code-ultimate-guide-mcp@1.3.1 and calling the MCP list methods. The repository contract comes from the generated product manifest and its live contract test. The Registry status was read from its public API after publication.

Use the MCP server when you want a coding client to search the guide, open the exact source section, retrieve a production template, or inspect Claude Code releases without loading the full guide into context.

The server runs locally over stdio. Version 1.3.1 has no first-party telemetry. Some tools can contact GitHub or Anthropic when invoked and can write local cache files. Downloads are not users, active installations, sessions, or executions.

Install the currently published package:

Terminal window
claude mcp add --scope user claude-code-guide -- npx -y claude-code-ultimate-guide-mcp@1.3.1
codex mcp add claude-code-guide -- npx -y claude-code-ultimate-guide-mcp@1.3.1

The guide is too large to inject in full for every question. The MCP server provides a smaller discovery and retrieval loop:

  1. search the structured guide index;
  2. return matching paths and anchors;
  3. read only the relevant source section;
  4. fetch a specific template or release record when needed.

This reduces irrelevant context. It does not guarantee that an answer is correct, current, or appropriate for a specific repository. The client still has to inspect the retrieved source and verify time-sensitive claims.

All examples below use the public npm version observed on 2026-09-01. Pinning the version makes installation reproducible.

User scope:

Terminal window
claude mcp add --scope user claude-code-guide -- npx -y claude-code-ultimate-guide-mcp@1.3.1
claude mcp list

Project scope in .mcp.json:

{
"mcpServers": {
"claude-code-guide": {
"type": "stdio",
"command": "npx",
"args": ["-y", "claude-code-ultimate-guide-mcp@1.3.1"]
}
}
}

Remove the user-scoped entry:

Terminal window
claude mcp remove --scope user claude-code-guide
Terminal window
codex mcp add claude-code-guide -- npx -y claude-code-ultimate-guide-mcp@1.3.1
codex mcp list

Remove it with:

Terminal window
codex mcp remove claude-code-guide

Add this entry to .cursor/mcp.json:

{
"mcpServers": {
"claude-code-guide": {
"type": "stdio",
"command": "npx",
"args": ["-y", "claude-code-ultimate-guide-mcp@1.3.1"]
}
}
}

Add this entry to .vscode/mcp.json:

{
"servers": {
"claude-code-guide": {
"type": "stdio",
"command": "npx",
"args": ["-y", "claude-code-ultimate-guide-mcp@1.3.1"]
}
}
}

Ask the client:

Search the Claude Code Ultimate Guide for session-scoped hooks. Open the most relevant source section and cite its path and heading.

The expected tool sequence is:

search_guide({ query: "session-scoped hooks" })
read_section({ path: "<path returned by search_guide>" })

For a template, start with search_examples, then call get_example with the returned name. For official documentation changes, run init_official_docs once, then refresh_official_docs and diff_official_docs when you need a new comparison.

MCP client
-> local stdio process started through npx
-> bundled indexes and datasets
-> optional GitHub fetch for guide content and templates
-> optional Anthropic fetch for official documentation snapshots
-> local cache under ~/.cache/claude-code-guide/

The MCP protocol travels through the child process standard input and standard output. There is no hosted guide MCP endpoint in this architecture. The client controls when the local process starts and which tools it calls.

The published 1.3.1 package exposes the following generated contract:

CapabilityCountPurpose
Tools17Search, source retrieval, templates, releases, threats, and official documentation snapshots
Resources6Reference index, release history, llms.txt, Agent Harness Map, distribution channels, and translation metadata
Prompts1claude-code-expert search and retrieval workflow

Tools by workflow:

WorkflowTools
Guide discoverysearch_guide, read_section, list_topics
Templatesget_example, list_examples, search_examples
Guide and release historyget_changelog, get_digest, get_release, compare_versions, get_cheatsheet
Security referenceget_threat, list_threats
Official Anthropic docsinit_official_docs, refresh_official_docs, diff_official_docs, search_official_docs

Resources:

  • claude-code-guide://reference
  • claude-code-guide://releases
  • claude-code-guide://llms
  • claude-code-guide://agent-harnesses
  • claude-code-guide://distribution-channels
  • claude-code-guide://translations

The published package searches 1,798 generated index entries. The public runtime snapshot verifies the capability names and counts without storing descriptions, arguments, content, local paths, or user data.

The package bundles the structured reference, release history, guide navigation, and other declared resource datasets. Initialization, list operations, and guide search use bundled data.

These tools can fetch content from GitHub when a local checkout or usable cache entry is not available:

  • read_section
  • get_example
  • get_cheatsheet
  • get_changelog
  • get_digest
  • get_threat
  • list_threats

Successful GitHub responses can be cached under ~/.cache/claude-code-guide/<version>/ for 24 hours. When GUIDE_ROOT points to a local guide checkout, content tools read from that checkout instead.

init_official_docs and refresh_official_docs fetch Anthropic documentation and store separate local snapshots. diff_official_docs and search_official_docs read those snapshots without making a network request.

Version 1.3.1 has no first-party telemetry, analytics endpoint, device identifier, or event upload. MCP requests and responses are not sent to this project’s author.

Network access is tool-specific:

OperationNetworkLocal write
List capabilities and search bundled indexNoNo
Read uncached guide content or templatesGitHub, when needed24-hour cache
Initialize or refresh official docsAnthropicLocal snapshots
Diff or search existing official-doc snapshotsNoNo

The MCP client, npm, npx, GitHub, Anthropic, and the machine operator can have their own logs or policies. Those systems are outside this server’s first-party telemetry boundary.

Centralized product telemetry is deliberately not implemented in 1.3.1. Download counts do not justify adding a tracking endpoint, and they cannot establish active usage.

Any future proposal requires a reviewed privacy decision before code exists. The minimum acceptance criteria are:

  • disabled by default and enabled by an explicit user action;
  • no query text, arguments, file paths, returned content, repository identity, hostname, or account identity;
  • a locally generated random identifier that the user can reset;
  • only package version, tool name, success or failure, and coarse latency;
  • a documented endpoint, retention period, deletion process, and exact payload schema;
  • an immediate disable mechanism and a public way to inspect emitted events.

Until those conditions are approved and implemented, the accurate claim is no first-party telemetry.

The server is partially offline, not fully offline.

  • capability listing, topic listing, release data, resources, and index search work from bundled content;
  • cached guide content can be read while the cache remains usable;
  • uncached Markdown sections and templates can require GitHub;
  • official-doc search and diff require snapshots previously created with network access;
  • the first npx execution can require npm registry access unless the package is already cached or installed.

For development or controlled offline use, install dependencies ahead of time, build the package, and start dist/index.js from a local checkout with GUIDE_ROOT set.

RequirementPublished 1.3.1
Transportlocal stdio
Node.js>=18.14.1
MCP SDK@modelcontextprotocol/sdk 1.30.0
Tested client configuration shapesClaude Code, Codex, Cursor, VS Code
Package manager entry pointnpx

Other MCP clients can work if they support local stdio servers and the same command-plus-arguments model. That statement is protocol compatibility, not a claim that every client release has been tested.

  • The full guide Markdown is not bundled.
  • Version 1.3.1 derives the MCP handshake from package.json and tests that contract.
  • The official Registry identity is case-sensitive: io.github.FlorianBruniaux/claude-code-guide.
  • Companion /ccguide:* command files live in the repository and are not installed by the npm package.
  • Search retrieves candidate sections; it does not replace source review or freshness checks.
  • No metric in this repository measures active users.
SymptomCheckAction
Client cannot start the servernode --version and npx --versionUse Node.js >=18.14.1 and verify npm registry access
Client shows no toolsclaude mcp list or codex mcp listCheck the server name, command, arguments, and configuration scope
Requested version cannot be downloadednpm view claude-code-ultimate-guide-mcp versionUse the exact version returned by npm and update the pinned client command
Source read fails offlineInspect ~/.cache/claude-code-guide/Retry with network access, set GUIDE_ROOT, or prefill the cache through an earlier successful read
Official-doc search says no snapshot existsCheck the official-doc cache directoryRun init_official_docs, which requires Anthropic network access
Configuration is staleRemove and re-add the serverUse the client-specific removal command shown above

Inspect the JSON-RPC surface with the MCP Inspector:

Terminal window
npx -y @modelcontextprotocol/inspector npx -y claude-code-ultimate-guide-mcp@1.3.1

Two machine-readable snapshots serve different questions:

  • mcp-stats.json records npm package history, daily distribution, and official-registry presence.
  • mcp-dashboard.json records one completed calendar month across npm, GSC, and GA4 with explicit availability states.

Snapshot time: 2026-08-31T22:41:52Z. Last complete UTC day: 2026-08-30. Unit: npm package downloads.

PeriodDownloads
Since launch, 2026-02-28 through 2026-08-308,590
Year to date, 2026-01-01 through 2026-08-308,590
Trailing 30 complete days, 2026-08-01 through 2026-08-305,252
Trailing 7 complete days, 2026-08-24 through 2026-08-30125

For the trailing 30 complete days, the daily median was 31.5, the mean was 175.1, the maximum was 727, and the median absolute deviation was 31.5. The mean is strongly affected by recorded spikes. The snapshot preserves those observations but does not assign a cause.

The completed July 2026 UTC snapshot records 1,459 npm downloads. GSC for the landing page, GSC for the portfolio page, and GA4 for the portfolio page are marked unavailable because the required Google access and property configuration were absent during collection. They are not reported as zero.

The monthly workflow uses exact page URL or path filters, publishes aggregate page and package metrics only, and keeps missing sources explicit. npm downloads can include installs, CI runs, cache misses, bots, and repeated downloads. They do not identify unique or active people.

Release changes must update the generated product manifest, pass npm run release:check in mcp-server/, and keep the public-package snapshot separate from repository candidates whenever their versions differ.