How to Set Up MCP Servers for Documentation Lookup in Claude Code: A Complete Guide

Question

Learn how to set up MCP servers for documentation lookup in Claude Code. Configure servers, set environment variables, and manage your context window for efficient code assistance.

Accepted Answer

To set up MCP servers for documentation lookup in Claude Code, copy server definitions from mcp-configs/mcp-servers.json into your global ~/.claude.json configuration, provide required environment variables, and use the disabledMcpServers array to disable unused servers and preserve your context window.

The everything-claude-code repository by affaan-m provides the reference implementation for extending Claude Code with live data through the Model Context Protocol (MCP). Setting up MCP servers for documentation lookup and other integrations involves configuring JSON definitions that specify how Claude invokes external tools like Context7 for real-time library documentation.

Global User Configuration (~/.claude.json)

Claude Code reads MCP server definitions from a global configuration file located at ~/.claude.json in your home directory. According to the source code in mcp-configs/mcp-servers.json, this file contains a map of server definitions that Claude can invoke to fetch external data.

To enable servers, copy the desired definitions into your global config:

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"],
      "description": "Live documentation lookup — use with /docs command and documentation-lookup skill."
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "YOUR_GITHUB_PAT_HERE" },
      "description": "GitHub operations – PRs, issues, repos"
    },
    "supabase": {
      "command": "npx",
      "args": ["-y", "@supabase/mcp-server-supabase@latest", "--project-ref=YOUR_PROJECT_REF"],
      "description": "Supabase database operations"
    }
  },
  "disabledMcpServers": []
}

Only servers listed under mcpServers become available in Claude Code sessions. Replace placeholder values with actual credentials and never commit this file to version control.

Project-Level Configuration (.claude.json)

Projects can fine-tune available servers by creating a .claude.json file in the repository root. This file inherits global definitions but allows you to override settings or disable specific servers.

The disabledMcpServers array, documented in mcp-configs/mcp-servers.json (lines 48-51), lets you exclude heavy or irrelevant servers:

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"],
      "description": "Documentation lookup for this project"
    }
  },
  "disabledMcpServers": [
    "clickhouse",
    "browser-use",
    "devfleet"
  ]
}

The repository recommends keeping under 10 MCPs enabled to preserve Claude's context window and token budget.

Enabling Documentation Lookup with Context7

The documentation-lookup skill provides live library documentation by wrapping the Context7 MCP server. Implemented in skills/documentation-lookup/SKILL.md, this skill orchestrates two key operations when you ask library-specific questions:

resolve-library-id: Queries Context7 with the library name and full user query to find the best matching library ID based on name match, benchmark score, reputation, and version.
query-docs: Fetches specific documentation content using the resolved library ID.

The skill caps the number of calls to 3 per question to minimize context usage (see line 55 of the skill documentation). It also explicitly redacts potential secrets from user prompts before sending queries to third-party services (line 90).

Configuring Additional MCP Integrations

Beyond documentation, you can enable various integrations by adding their definitions to ~/.claude.json:

GitHub: Requires GITHUB_PERSONAL_ACCESS_TOKEN environment variable for PR and issue automation.
Supabase: Requires --project-ref argument matching your Supabase project reference.
Browser-use, ClickHouse, Fal-ai: Specialized servers for web automation, analytics, and AI model inference.

Each server definition follows the same structure: a command (usually npx), args array, optional env variables, and a description field explaining the server's purpose.

Adding Custom HTTP-Based MCP Servers

For internal APIs or services that expose HTTP endpoints, define servers with type: "http" instead of command-based execution:

"my-custom-api": {
  "type": "http",
  "url": "https://api.myservice.com/mcp",
  "description": "Custom internal API for proprietary data"
}

Command-based servers require command and args arrays, while HTTP servers need only url and type specifications.

Security and Performance Best Practices

Follow these guidelines when setting up MCP servers:

Never commit secrets: Store API keys only in ~/.claude.json outside version control.
Limit active servers: Keep under 10 enabled MCPs to preserve context window, as warned in mcp-configs/mcp-servers.json.
Version-specific queries: Include version numbers in questions to help resolve-library-id select accurate documentation.
Redact sensitive data: The documentation-lookup skill explicitly scrubs potential secrets from prompts before querying external services.

Summary

MCP server configurations reside in mcp-configs/mcp-servers.json and must be copied to your global ~/.claude.json file to activate them.
Project-level overrides use .claude.json in the repository root with the disabledMcpServers array to disable unnecessary servers.
Context7 powers documentation lookup through the documentation-lookup skill, which calls resolve-library-id and query-docs methods defined in skills/documentation-lookup/SKILL.md.
Environment variables such as GITHUB_PERSONAL_ACCESS_TOKEN provide authentication for services like GitHub and Supabase.
Performance constraints require limiting active servers to under 10 to maintain Claude's context window efficiency.

Frequently Asked Questions

What file do I edit to enable MCP servers globally in Claude Code?

Create or edit ~/.claude.json in your home directory and add server definitions from the repository's mcp-configs/mcp-servers.json file. This global configuration makes servers available across all Claude Code sessions on your machine.

How does the documentation-lookup skill fetch library information?

The skill calls the Context7 MCP server's resolve-library-id method to identify the correct library based on name, version, and relevance scoring, then uses query-docs to retrieve specific documentation content. It limits itself to 3 calls per question to conserve context window.

Why should I disable unused MCP servers?

Each active MCP server consumes tokens from Claude's context window. The repository recommends disabling servers via the disabledMcpServers array to stay under the 10-server limit, ensuring optimal performance and preventing context overflow during complex operations.

Can I use HTTP-based MCP servers instead of command-based ones?

Yes, define servers with "type": "http" and provide a url field pointing to your MCP endpoint. This avoids spawning npm processes and works well for proprietary internal APIs that expose Model Context Protocol endpoints over HTTP.