Claude Code MCP Directory

A practical Claude Code MCP directory for choosing GitHub MCP, Context7 MCP, Playwright MCP, transports, scopes, authentication, security, and team rollout patterns.

MCP lets Claude Code connect to external tools, documentation, browser sessions, repositories, databases, monitoring systems, and internal APIs. The useful question is not "how many MCP servers can I install?" The useful question is: which server removes a repeated workflow bottleneck without exposing more data or write access than the task needs?

Last checked on May 24, 2026. Claude Code MCP supports local stdio servers, remote HTTP servers, older SSE servers, OAuth flows, local/project/user scopes, Claude.ai connectors, and enterprise-managed MCP policy. HTTP is now the preferred remote transport where available; SSE is deprecated in the current Claude Code docs.

Quick Answer

Start with one MCP server only when the workflow repeatedly needs context outside the repository.

Use GitHub MCP when Claude needs issue, PR, review, release, or repository context.
Use Context7 MCP when Claude needs current library and framework documentation before editing code.
Use Playwright MCP when Claude needs to inspect a local page, interact with UI, or capture screenshot evidence.
Use the MCP security checklist before adding write scopes, tokens, authenticated browser sessions, databases, or internal systems.
Use local or project configuration only when you understand who owns credentials, updates, logs, and incident response.

If a shell command, copied URL, or static CLAUDE.md note solves the problem, do not add MCP yet.

Choose An MCP Server By Workflow

Workflow need	Best entry point	First safe task	Main risk
Read issue or pull request context	GitHub MCP	Summarize an issue or inspect a PR before implementation.	Write scopes, repo visibility, accidental comments or changes.
Check current docs before editing code	Context7 MCP	Fetch docs for one library and verify one API decision.	Trusting stale or irrelevant docs without source checking.
Verify UI behavior in a browser	Playwright MCP	Open a local route, click through one flow, capture a screenshot.	Authenticated sessions, production actions, flaky environment state.
Review MCP permissions before rollout	MCP security checklist	Map read/write tools, secrets, logs, and owner before install.	Broad access without team governance.
Fix a server that will not load	MCP not connecting	Use `/mcp` and inspect config, startup, auth, and scopes.	Changing multiple variables at once.
Learn the basic setup flow	MCP server tutorial	Add one read-only server and run one focused prompt.	Installing servers because they look useful, not because a workflow needs them.

Good MCP adoption is usually boring: one server, one task, read-only first, measurable improvement.

What MCP Adds To Claude Code

Model Context Protocol gives Claude Code a structured way to call tools outside the chat. A server can expose tools, resources, and prompts, while Claude Code remains the client that decides what context or action to request.

MCP is useful when:

the needed context lives in another system, such as issues, PRs, browser state, docs, monitoring, tickets, or internal APIs;
the task repeats often enough that a tool call is safer than copy-paste;
the data source changes too often for static project memory;
the team can define clear read/write boundaries;
the value is visible in a real workflow, not just in a demo.

MCP is not useful when the workflow is vague, the server requires broad credentials, or nobody owns the server after installation.

Choose The Transport Deliberately

Claude Code can connect to local stdio servers and remote servers. Pick the transport that matches ownership and risk, not the newest-looking option.

Transport	Use when	Watch out for
HTTP / Streamable HTTP	The server is hosted, cloud-based, organization-managed, or uses OAuth.	Network trust, auth scopes, endpoint governance, logs, and token handling.
stdio	The server runs locally, wraps a CLI, uses Docker, or needs direct machine/project access.	Inherits local environment, can touch local files, and may depend on machine-specific paths.
SSE	The server only supports older remote MCP deployments.	Claude Code docs now mark SSE as deprecated; prefer HTTP where available.

For remote production systems, HTTP with proper OAuth or controlled headers is usually easier to govern than a local command with hidden credentials. For local developer tools, stdio can be simpler and safer if credentials stay local.

Choose The Scope Deliberately

MCP configuration scope changes who sees the server and where configuration is stored.

Scope	Loads in	Shared with team	Best use
Local	Current project only	No	Personal experiments, private credentials, first trials.
Project	Current project only	Yes, through `.mcp.json`	Team-approved servers and repeatable repository workflows.
User	All your projects	No	Personal cross-project tools that do not belong in one repo.
Managed / enterprise	Admin-controlled machines	Admin-defined	Organization policy, allowlists, denylists, and mandatory servers.

Do not put secrets into versioned project config. Project scope is for server definitions and shared intent, not raw credentials.

Authentication And OAuth

Many remote MCP servers require authentication. In Claude Code, /mcp is the place to inspect status and complete authentication flows.

Practical rules:

Use /mcp after adding a server and complete the browser auth flow when prompted.
Use /status if Claude.ai connectors do not appear; active API keys or third-party providers can change which connectors load.
Use fixed OAuth callback ports only when the server requires a pre-registered redirect URI.
Use pinned OAuth scopes when the server advertises more access than your workflow needs.
Use dynamic headers or helper scripts only when your team can audit and maintain them.
Never commit tokens, client secrets, or personal credentials.

If a remote server returns auth errors, do not keep reinstalling it. Inspect scopes, callback URL, account identity, and whether the server expects dynamic client registration or pre-configured credentials.

Claude.ai Connectors And Claude Code

Claude.ai connectors use the same MCP infrastructure. If you sign in to Claude Code with a Claude.ai subscription account, connectors added in Claude.ai can appear inside Claude Code.

This is convenient, but it can surprise teams:

connectors may not load when an API key, auth token helper, Bedrock, Vertex, or another provider is active;
a server added directly in Claude Code can take precedence over a connector that points to the same URL;
Team and Enterprise environments may restrict who can add connectors;
/mcp and /status are the fastest way to see what Claude Code actually loaded.

For production work, document whether the workflow expects Claude.ai connectors, Claude Code local config, project .mcp.json, or managed enterprise config.

Safe Rollout Flow

Use this order for a new MCP server:

Write the workflow in one sentence: "Use GitHub MCP to read issue context before implementation."
Choose one server and one task: avoid installing a bundle of servers at once.
Start read-only: use the smallest permissions or OAuth scopes that prove value.
Add locally first: keep it private until the workflow is validated.
Run a real task: capture what became faster, safer, or more accurate.
Document boundaries in CLAUDE.md: purpose, allowed actions, owner, and what not to do.
Promote to project scope only after review: use .mcp.json for team-approved config.
Add hooks or allowlists when risk grows: especially for write tools and internal systems.

The first successful MCP setup should feel narrow. Broad access comes later, if ever.

Setup Pattern

The exact command depends on the server, but the shape is usually:

claude mcp add --transport http <server-name> <server-url>

For a remote server with a bearer token:

claude mcp add --transport http secure-api https://api.example.com/mcp \
  --header "Authorization: Bearer <token>"

For a local stdio server:

claude mcp add --transport stdio <server-name> -- <command> [args...]

After adding a server:

claude mcp list
claude mcp get <server-name>

Then open Claude Code and run:

/mcp

Start with a read-only prompt such as "list the tools this server exposes" or "retrieve one issue title" before allowing edits, comments, writes, browser actions, or database access.

What Not To Add Yet

Skip an MCP server when:

a simple CLI command already solves the problem;
the server needs broad credentials but the use case is vague;
nobody owns updates, token rotation, incident response, or logs;
the server reads production data for a task that can be tested locally;
the team cannot explain what the server can read or write;
the same information belongs in CLAUDE.md, a tutorial, or a one-time prompt;
the first workflow is "maybe useful someday."

MCP should reduce repeated friction. It should not become a hidden permission layer nobody understands.

Team And Enterprise Governance

Teams should treat MCP like any other integration:

Governance question	Why it matters
Who owns the server?	Someone must update config, rotate credentials, and respond to incidents.
Is it read-only or write-capable?	Write tools need reviews, hooks, or explicit approval habits.
Where is it configured?	Local, project, user, Claude.ai connector, and managed config behave differently.
How are secrets handled?	Secrets should stay in keychains, provider auth flows, env vars, or audited helpers, not docs.
Can users add unapproved servers?	Enterprise allowlists and denylists may be needed for managed environments.
What is logged?	Tool calls can touch sensitive project, browser, or customer data.

For high-risk environments, review MCP security checklist, Tools allowlist, and Hooks before giving Claude Code write access through MCP.