Claude Code MCP in Practice: GitHub, CodeGraph & API complete architecture

What is MCP triple-connect?

Claude Code MCP uses the MCP protocol to bring GitHub (issues and intent), CodeGraph (repo structure and blast radius), and APIs (environment and ground truth) into Claude Code so the agent can see what to do, what to change, and whether it is right—we call that triple-connect.

This page is the Cloud Mac AI Stack overview for MCP: one diagram for the big picture, then links to the setup guide, architecture article, or CodeGraph walkthrough. Commands and troubleshooting stay in those articles. The previous L2 post: parallel memory scheduling.

Why several articles instead of one long post

Readers arrive in different modes: some need the environment working now; others want to know why GitHub and CodeGraph belong together; others care most about tokens and permissions. Packing install steps, theory, and permission checklists into one URL forces people hunting for JSON to scroll past long mcpServers blocks, and interrupts anyone who only wants the map first.

So we split by task: this page shows what each piece does in one diagram; open the setup guide when you are ready to wire things up; read the architecture article when you need to explain the design to your team. Treat this as the front door—pick the next article from here.

What this page does in the series

Think of this as the table of contents for MCP triple-connect: see how the pieces fit, then open the article that matches your job. It does not run installs for you, and it is not a substitute for the CodeGraph or Runner guides.

This page does not include:

• Step-by-step Claude Code MCP install (no copy-paste command walkthrough)
• Long architecture derivations or Context Engineering theory
• Deploying or operating CodeGraph on its own (see CodeGraph in five minutes)

This page does help you pick the next read based on your goal or symptom—setup, architecture, graph, Runner, runtime, or permissions.

Not here—open the right article

If you need what is in the left column, use “Go to” instead of hunting for steps on this overview.

Quick navigation: what to read next

Pick one topic that matches your goal; scenario shortcuts and symptom tables below go deeper.

Need green CI checks after a push? See GitHub Runner guide. Lid closed and MCP drops or the index goes stale? See Cloud Mac vs local Mac

What are you trying to do?

People search with different keywords, but goals usually fall into four buckets. Pick yours to save time.

① Get the environment running

People often search for: Claude Code MCP setup · install MCP server · configure GitHub MCP · connect Claude Code to GitHub

Your goal: Run the toolchain and see mcp__* tools in Claude Code.
Suggested read Setup guide

② Understand the overall design

People often search for: MCP architecture explained · Context Diff Fact · why MCP needed

Your goal: Understand the layers and align your team.
Suggested read Architecture article

③ Repo graph / impact analysis only

People often search for: CodeGraph MCP · codegraph_impact · repo graph Claude Code

Your goal: Wire the graph first; GitHub issue flow can wait.
Suggested read CodeGraph in five minutes; for full triple-connect later see Setup guide

④ Permissions and safety

People often search for: MCP permissions · token scope · read-only API MCP

Your goal: Control token scope, isolate writable tools, keep staging read-only.
Suggested read Permissions guide; for a minimal working setup first, follow the setup guide, then tighten permissions

Still unsure? Use quick navigation or the problem lookup table below.

Common scenarios—start here

Same triple-connect, different entry points—choose the line that sounds like you.

• First time with MCP: skim this diagram → setup guide
• CodeGraph is already running; add GitHub → setup guide (GitHub section); graph stays in CodeGraph in five minutes
• CI: green checks after push → GitHub Runner guide; how MCP and Runner differ → architecture article
• Tightening tokens, env vars, read-only API → Permissions guide; interim minimal connect via setup guide
• Installed but tools missing or flaky → setup guide · troubleshooting
• Architecture only—no install yet → architecture article

Something wrong? Match your symptom

Symptoms below point to the article that usually helps. Root-cause detail and commands live there—not on this overview.

Three common reading paths

First contact with MCP

Open Claude Code → read this overview (triple-connect) → follow the setup guide for mcpServers and confirm tools in chat → add the GitHub Runner guide when you want checks after push.

Graph is done—add GitHub only

Finished CodeGraph in five minutes → return here to place GitHub in the map → wire GitHub MCP in the setup guide → optional architecture article for boundaries.

Team rollout: AI-assisted CI

This overview → setup guide (triple-connect) → Runner guide (automated checks) → architecture article (team alignment) → Permissions guide to tighten tokens.

One diagram, clear roles

GitHub      = intent
CodeGraph   = context
API         = truth
Claude Code = diff
Runner      = fact

FAQ

Bookmark this page as your MCP triple-connect starting point. English slug: claude-code-mcp-github-files-api-chain.