MCP Servers

Claude Kit includes optional Model Context Protocol (MCP) server configurations that extend Claude’s capabilities with external tools.

How MCP Works

MCP servers give Claude access to tools it doesn’t have natively — browser automation, persistent memory, real-time documentation, and structured reasoning. They run as local processes that Claude communicates with during your session.

MCP servers are configured via /claudekit:init, which adds them to your project’s .mcp.json with automatic platform detection (Windows vs macOS/Linux).

Available Servers

Context7

Purpose: Real-time library documentation lookup.

Fetches current documentation for any library, framework, or API. Use instead of relying on Claude’s training data, which may be outdated.

You: "How do I set up middleware in Next.js 15?"

Claude fetches current Next.js 15 docs via Context7
→ Answers with up-to-date API syntax

Best for: API syntax, configuration, version migration, library-specific debugging.

Setup: Run /claudekit:init and select Context7.

Sequential Thinking

Purpose: Structured step-by-step reasoning with explicit thought chains.

Provides a tool for multi-step analysis where each step has a confidence score and the chain can revise earlier steps as new evidence comes in.

Investigation:
  Step 1: Capture the error → confidence: 0.9
  Step 2: Form hypothesis (X causes Y when Z) → confidence: 0.7
  Step 3: Test hypothesis with instrumentation → confidence: 0.85
  Step 4: Verify the fix doesn't regress → confidence: 0.95

Best for: Complex debugging, architectural trade-off analysis, security review where multiple hypotheses need to be tracked simultaneously.

Setup: Run /claudekit:init and select Sequential Thinking.

Memory

Purpose: Persistent knowledge graph across sessions.

Stores entities, relationships, and observations that persist across conversations. Claude can recall project decisions, user preferences, and architectural context the next time you sit down.

Session 1: "We decided to use PostgreSQL RLS for multi-tenancy"
  → Stored as entity + decision observation

Session 2 (a week later): "What did we decide about multi-tenancy?"
  → Retrieved from memory graph

Best for: Long-running projects, decision tracking, building up codebase knowledge over time.

Setup: Run /claudekit:init and select Memory.

Filesystem

Purpose: Sandboxed file operations with configurable allowed directories.

Useful for projects with strict file access requirements (e.g., when you want Claude restricted to a specific subtree of a monorepo, or when you’re operating in a regulated environment with audit-trail requirements).

Best for: Projects with strict file access requirements; regulated codebases.

Setup: Run /claudekit:init and select Filesystem.

Playwright

Purpose: Browser automation for testing and verification.

Enables Claude to control a real browser for E2E testing, visual verification, and runtime UI checks.

You: "Verify the login flow works in production"

Claude launches a browser via Playwright MCP:
  → Navigate to /login
  → Fill email and password
  → Click submit
  → Verify redirect to /dashboard
  → Take screenshot for evidence

Best for: E2E testing, visual regression checks, the non-IDE verification step in verification-gate.

Setup: Run /claudekit:init and select Playwright.

Setup

Prerequisites

MCP servers require Node.js installed on your system.

Enabling Servers

Run the setup wizard and select which servers to configure:

/claudekit:init

Or install all servers at once:

/claudekit:init --all

The wizard automatically detects your platform and configures the correct command format in .mcp.json. Restart Claude Code after configuration.

Which skills benefit from each server

MCP Server	Skills that get the most lift
Context7	`audit-dependencies` (verify advisories against current docs), `investigate-root-cause` (confirm framework behavior matches docs), `shape-spec` (research library options before committing), `incremental-shipping` (read changelog before bumping a dep)
Sequential Thinking	`investigate-root-cause` (the 4-phase loop benefits from explicit confidence tracking), `plan-review-architecture` (multi-dimensional scoring), `shape-spec` (working through alternatives systematically)
Memory	`shape-spec` (recall design decisions across sessions), `map-codebase` (build up codebase knowledge over time), `release-and-changelog` (recall release history)
Playwright	`test-first` (E2E test cases for UI flows), `verification-gate` (the non-IDE verification step — exercising the change in a real browser)
Filesystem	Project-wide; no specific skill mapping. Use when you need scoped file access.

MCP servers are optional — claudekit’s spine works without them. They add capability where they fit; the skills enforce discipline regardless.