MCP Config Generator (.mcp.json)

Model Context Protocol (MCP): how it works

The Model Context Protocol (MCP) is an open standard that lets AI tools call external servers for data and actions, such as GitHub issues, a Postgres database or files on disk. Each client lists its servers in a JSON file under an mcpServers key: a command that starts a local server (with args and env), or a URL for a remote one. Claude Code reads a project's .mcp.json from the repo root and asks before using its servers. Claude Desktop uses claude_desktop_config.json, and Cursor uses .cursor/mcp.json.

Where the file goes

./.mcp.jsonClaude Code, projectCommitted with the repo so the team shares the same servers. Claude Code asks for approval before using them.
~/Library/Application Support/Claude/claude_desktop_config.jsonClaude DesktopOn Windows: %APPDATA%\Claude\claude_desktop_config.json. Restart Claude Desktop after editing.
.cursor/mcp.jsonCursor, projectUse ~/.cursor/mcp.json for servers available in every project.

What goes in it

Local serverA command the client starts on your machine. Give a filesystem server only the folders it needs.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "./src", "./docs"]
    }
  }
}

Remote server with a tokenClaude Code expands ${VAR} from your environment when it loads .mcp.json, so the token never lands in git.

{
  "mcpServers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": { "Authorization": "Bearer ${GITHUB_PERSONAL_ACCESS_TOKEN}" }
    }
  }
}

Tips

Never commit tokens. Use ${VAR} references in .mcp.json and set the values in your shell.
Give filesystem servers only the folders they need, never your home directory or a drive root.
Use read-only database users and fine-grained, read-only GitHub tokens unless the agent has to write.
Only add servers you trust: a server sees what you send it and acts with the credentials you give it.
Restart the client after editing the config; most read it only at startup.

MCP Config presets (9)

Each one opens the generator with that file filled in for the stack. Change what you need, then copy or download it.

Which file do I need?

Each AI coding tool reads its own file from its own place. Find your tool, put the file where the table says, and check when it actually gets loaded. Every file name opens a generator for that format.

AI instruction files by tool, with location and load timing
AI tool	File	Where it goes	When it's loaded
Agent rules and tool config
Cursor	<rule-name>.mdc	Goes in: `.cursor/rules/`	Loaded: Per rule: always, when `globs` match the files in play, when the agent finds the `description` relevant, or on @-mention.
Claude Code	SKILL.md	Goes in: `.claude/skills/<skill-name>/`, or `~/.claude/skills/` for all projects	Loaded: Name and description at startup. The full file only when a task matches or you run `/skill-name`.
Claude Code	CLAUDE.md	Goes in: Repo root, or `~/.claude/CLAUDE.md` for all projects	Loaded: Every session, in full. A `CLAUDE.md` in a subfolder loads when Claude works there.
Codex, Cursor, Copilot coding agent and others	AGENTS.md	Goes in: Repo root. Nested files cover their subfolder.	Loaded: At the start of each task. For Claude Code, add `@AGENTS.md` to `CLAUDE.md`.
Claude Code, Claude Desktop, Cursor	.mcp.json	Goes in: Repo root for Claude Code. Claude Desktop uses `claude_desktop_config.json`; Cursor uses `.cursor/mcp.json`.	Loaded: At startup, so restart after editing. Claude Code asks before using servers from `.mcp.json`.
Windsurf	<rule-name>.md	Goes in: `.windsurf/rules/`	Loaded: Per rule `trigger`: always on, glob match, model decision, or manual @-mention.
GitHub Copilot	copilot-instructions.md	Goes in: `.github/`	Loaded: With every Copilot Chat, code review and coding agent request in the repo. Not used for inline completions.
ChatGPT / OpenAI API	openai-custom-instructions.md	Goes in: `prompts/`, as a copy you version	Loaded: Only when pasted into ChatGPT custom instructions or sent as the system message.
Gemini API / AI Studio	gemini-system-instructions.json	Goes in: `prompts/`, as a copy you version	Loaded: Only when passed as `systemInstruction` or pasted into AI Studio.
Ignore files
Cursor	.cursorignore	Goes in: Repo root	Loaded: Always. Indexing and AI features skip matching files (best effort: terminal commands can still read them).
Claude Code	.claudeignore	Goes in: Repo root	Loaded: Keep it as your pattern list, and mirror secret paths in `permissions.deny` in `.claude/settings.json`, the setting Claude Code documents for blocking reads.
Reference docs
Any LLM or crawler	llms.txt	Goes in: Repo root, or your site root as `/llms.txt`	Loaded: Only when a tool or person fetches it. Nothing loads it automatically.
Any agent	ARCHITECTURE.md	Goes in: Repo root	Loaded: When your rules file points to it, or when the agent explores the repo.
Planning docs
Any agent	PRD.md	Goes in: Repo root	Loaded: When your rules file tells the agent to read it.
Any agent	DESIGN.md	Goes in: Repo root	Loaded: When your rules file tells the agent to read it, usually before UI work.
Any agent	TASK.md	Goes in: Repo root	Loaded: At the start and end of a task, if your rules say to read and update it.
Any agent	MEMORY.md	Goes in: Repo root	Loaded: At session start, if your rules say so. Agents add lessons as they go.

What a good rule looks like

Agents follow rules they can check. Swap advice for commands, paths and explicit “never” lines.

Give commands, not wishes

CLAUDE.md or AGENTS.md

Vague

## Guidelines
- Write clean code.
- Follow best practices.
- Make sure tests pass.

Specific

## Commands
- Test one file: pnpm vitest run src/lib/date.test.ts
- Typecheck: pnpm tsc --noEmit
- Never run pnpm db:reset. It wipes the shared dev database.

Why it works: The agent can run each line as written, and the "never" rule names the exact command and the reason.

Scope rules to the files they cover

.cursor/rules/components.mdc

Vague

---
alwaysApply: true
---
Use React best practices.

Specific

---
globs: src/components/**/*.tsx
alwaysApply: false
---
- Reuse Button and Input from @/components/ui; never add a UI library.
- One component per file, named export, props typed with an interface.

Why it works: It loads only when a component file is in play, and every line is something the model can check.

Tell a skill when to run

SKILL.md

Vague

---
name: helper
description: Helps with code.
---

Specific

---
name: db-migration
description: Writes and reviews Prisma migrations. Use when schema.prisma changes or the user asks for a migration.
---

Why it works: Claude picks a skill by its description, so say what it does and when to use it.

Copy-ready starters

Complete files from the preset library. Copy one as it is, or open it to change the stack, commands and rules before you download it.

Next.js 15 App RouterCLAUDE.md
CLAUDE.md
Define standard commands: build, test, lint, and dev server.
Next.js 15 App RouterCursor Rules
.cursor/rules/nextjs-15.mdc
Enforce React Server Components (RSC) by default; restrict 'use client' strictly to interactive leaf components.
Automated Code ReviewerClaude Skill
.claude/skills/codebase-auditor/SKILL.md
Scope scanning to the repository and skip .env files, keys, and credential stores.
Multi-Agent System OrchestrationAGENTS.md
AGENTS.md
Define subagent roles: Lead Architect, Frontend Engineer, Backend Specialist, Security Reviewer.
Python FastAPI & Pydantic v2Cursor Rules
.cursor/rules/fastapi.mdc
Use strict Pydantic v2 BaseModels with Field validation for all input and output schemas.
React 19 & ViteCopilot
.github/copilot-instructions.md
Use React 19 action hooks (useActionState, useOptimistic, useFormStatus) for async mutations.
Tailwind CSS v4Cursor Rules
.cursor/rules/tailwind-v4.mdc
Configure themes directly in globals.css using @theme and @import 'tailwindcss';.
GitHub Operations MCPMCP Config
.mcp.json or claude_desktop_config.json
Uses official @modelcontextprotocol/server-github package via npx.

Browse the full preset library

Frequently asked questions

Where does the MCP config file go?

For Claude Code, .mcp.json in the repo root, or claude mcp add for personal servers. Claude Desktop uses claude_desktop_config.json (Settings > Developer > Edit Config). Cursor uses .cursor/mcp.json in the project or ~/.cursor/mcp.json for all projects.

Can I use the same config in VS Code?

Mostly. VS Code reads .vscode/mcp.json, where the top-level key is servers instead of mcpServers. The server entries have the same shape.

How do I keep API tokens out of the file?

Write ${GITHUB_PERSONAL_ACCESS_TOKEN} instead of the token. Claude Code expands environment variables in .mcp.json, so the real value stays in your shell or secret manager.

CLAUDE.md vs AGENTS.md: which do I need?

Claude Code reads CLAUDE.md. AGENTS.md is the shared file that OpenAI Codex, Cursor, GitHub Copilot's coding agent and a growing list of other agents read. If your team uses more than one tool, put the rules in AGENTS.md and add the line @AGENTS.md to CLAUDE.md so Claude Code imports the same file.

Why is Cursor ignoring my rules?

Usually it's the frontmatter. A rule with alwaysApply: false and no globs only loads when the agent decides its description is relevant, or when you @-mention it. Check that the file is in .cursor/rules/, that globs match the files you are editing, and that the description says when the rule applies. Rules don't affect Tab autocomplete at all.

How long should CLAUDE.md be?

Aim for under 200 lines. Claude Code loads the whole file into every session, so each line costs context and competes with the others for attention. Keep build and test commands, the folder layout and hard "never" rules in it. Move long procedures into skills or separate docs that the file points to.

Where do Claude Code skills go?

Each skill is a folder with a SKILL.md inside: .claude/skills/<skill-name>/SKILL.md for one project, or ~/.claude/skills/<skill-name>/SKILL.md for all of your projects. Claude Code keeps only each skill's name and description in context, and loads the full file when a task matches it or when you type /skill-name.

How do I stop AI tools from reading my .env files?

For Cursor, list secrets such as .env*, *.pem and *.key in .cursorignore at the repo root. For Claude Code, add deny rules like Read(./.env) and Read(./.env.*) under permissions.deny in .claude/settings.json, which is the setting Claude Code documents for blocking file reads. Neither replaces keeping secrets out of the repo.

Do AI tools read PRD.md, TASK.md or ARCHITECTURE.md automatically?

No. Agents load their own rules file (CLAUDE.md, AGENTS.md or Cursor rules) and nothing else on their own. Name the other docs in that file, for example "Read PRD.md and TASK.md before you start", and the agent opens them when it needs them. In CLAUDE.md you can also write @PRD.md, but that loads the file into every session.

Can it write rules from my existing code?

Yes. Click Read your code and point it at a GitHub repository or a local folder. Besides the manifest, it reads a sample of your source files and any existing CLAUDE.md, AGENTS.md or Cursor rules, then writes only the conventions most of the code follows, such as test layout, file naming, import alias, error classes and logger, each with how many files back it. The examples in the file quote a real function from your code.

Does AI Skill Studio upload my code?

No. Presets, code reading, editing and ZIP export all run in your browser, and nothing you type or import is sent to DevScratchpad. Reading a local folder happens entirely in your browser. Reading a GitHub repository is the only network call, and your browser sends it straight to GitHub.

MCP config generator: .mcp.json for Claude Code, Claude Desktop and Cursor

Target Standard & File Format

MCP Server

Target File Location

Model Context Protocol (MCP): how it works

Where the file goes

What goes in it

Tips

MCP Config presets (9)

Which file do I need?

What a good rule looks like

Give commands, not wishes

Scope rules to the files they cover

Tell a skill when to run

Copy-ready starters

Frequently asked questions

Where does the MCP config file go?

Can I use the same config in VS Code?

How do I keep API tokens out of the file?

CLAUDE.md vs AGENTS.md: which do I need?

Why is Cursor ignoring my rules?

How long should CLAUDE.md be?

Where do Claude Code skills go?

How do I stop AI tools from reading my .env files?

Do AI tools read PRD.md, TASK.md or ARCHITECTURE.md automatically?

Can it write rules from my existing code?

Does AI Skill Studio upload my code?