Which file does Claude Code read for MCP servers?
For a project, .mcp.json at the repository root, with servers under mcpServers. Remote entries need "type": "http"; an entry with a url and no type is read as a local command and skipped.
Choose the MCP servers your agent may use and get a working config for Claude Code, Cursor, VS Code and Codex from one list. 37 verified servers, 15 ready-made sets, and no secret ever written into a file.
{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"playwright": {
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
},
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"."
]
},
"memory": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
}
}
}
{
"mcpServers": {
"context7": {
"url": "https://mcp.context7.com/mcp"
},
"playwright": {
"type": "stdio",
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
},
"filesystem": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"."
]
},
"memory": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
}
}
}
{
"inputs": [],
"servers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"playwright": {
"type": "stdio",
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
},
"filesystem": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"${workspaceFolder}"
]
},
"memory": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
}
}
}
# Codex MCP servers (generated by 2prompt). Merge into ~/.codex/config.toml, or keep it as
# .codex/config.toml in the repository: Codex reads project config for trusted projects only.
# Secrets are never written here; env_vars and bearer_token_env_var name variables to read.
[mcp_servers.context7]
url = "https://mcp.context7.com/mcp"
[mcp_servers.playwright]
command = "npx"
args = ["@playwright/mcp@latest"]
[mcp_servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "--read-only", "."]
[mcp_servers.memory]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-memory"]
Commit the files for the clients your team uses. They hold no secrets: set the listed environment variables on each machine, or let VS Code prompt for them.
Fifteen server sets for common kinds of work. Each card writes all four client configs: load it to change the servers, or download the set as a zip.
A real browser to click through the UI, DevTools for console and network errors, current framework docs, Figma for design context and shadcn/ui components.
Preconfigured for: React, Next.js and Vite front-end work.
.mcp.json{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"playwright": {
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
},
"chrome-devtools": {
"command": "npx",
"args": [
"-y",
"chrome-devtools-mcp@latest",
"--no-usage-statistics"
]
},
"shadcn": {
"command": "npx",
"args": [
"shadcn@latest",
"mcp"
]
},
"figma": {
"type": "http",
"url": "https://mcp.figma.com/mcp"
}
}
}
Postgres in restricted (read-only) mode, an in-memory DuckDB for local analysis, URL fetching, time-zone math and access to one data folder.
Preconfigured for: Analysts and data engineers working in Python notebooks and scripts.
.mcp.json{
"mcpServers": {
"postgres": {
"command": "uvx",
"args": [
"postgres-mcp",
"--access-mode=restricted"
],
"env": {
"DATABASE_URI": "${DATABASE_URI}"
}
},
"duckdb": {
"command": "uvx",
"args": [
"mcp-server-motherduck",
"--db-path",
":memory:",
"--read-write"
]
},
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"./data"
]
},
"fetch": {
"command": "uvx",
"args": [
"mcp-server-fetch"
]
},
"time": {
"command": "uvx",
"args": [
"mcp-server-time"
]
}
}
}
Grounds the agent in current documentation instead of memory: Context7 for libraries, Microsoft Learn, AWS and Cloudflare docs, Exa web search and DeepWiki for GitHub repositories.
Preconfigured for: Any team that wants fewer hallucinated APIs.
.mcp.json{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"microsoft-learn": {
"type": "http",
"url": "https://learn.microsoft.com/api/mcp"
},
"aws-docs": {
"command": "uvx",
"args": [
"awslabs.aws-documentation-mcp-server@latest"
],
"env": {
"FASTMCP_LOG_LEVEL": "ERROR"
}
},
"cloudflare-docs": {
"type": "http",
"url": "https://docs.mcp.cloudflare.com/mcp"
},
"deepwiki": {
"type": "http",
"url": "https://mcp.deepwiki.com/mcp"
},
"exa": {
"type": "http",
"url": "https://mcp.exa.ai/mcp"
}
}
}
Next.js DevTools for build and runtime errors, a Supabase server scoped to one project in read-only mode, Playwright for end-to-end checks and Context7 docs.
Preconfigured for: Next.js App Router apps on Supabase.
.mcp.json{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"playwright": {
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
},
"next-devtools": {
"command": "npx",
"args": [
"-y",
"next-devtools-mcp@latest"
]
},
"supabase-local": {
"command": "npx",
"args": [
"-y",
"@supabase/mcp-server-supabase@latest",
"--read-only",
"--project-ref=your-project-ref"
],
"env": {
"SUPABASE_ACCESS_TOKEN": "${SUPABASE_ACCESS_TOKEN}"
}
}
}
}
GitHub issues and pull requests through the official remote server, local git for history and diffs, and filesystem access to the working tree.
Preconfigured for: Developers who live in issues and PRs.
.mcp.json{
"mcpServers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": {
"Authorization": "Bearer ${GITHUB_PAT}"
}
},
"git": {
"command": "uvx",
"args": [
"mcp-server-git",
"--repository",
"."
]
},
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"."
]
}
}
}
Linear for issues, Notion for specs, Figma for designs and GitHub for code: the agent can read the ticket, the spec and the design before it writes anything.
Preconfigured for: Product engineers working from tickets and specs.
.mcp.json{
"mcpServers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": {
"Authorization": "Bearer ${GITHUB_PAT}"
}
},
"linear": {
"type": "http",
"url": "https://mcp.linear.app/mcp"
},
"notion": {
"type": "http",
"url": "https://mcp.notion.com/mcp"
},
"figma": {
"type": "http",
"url": "https://mcp.figma.com/mcp"
}
}
}
Sentry errors and traces, PostHog analytics and flags, and Chrome DevTools to reproduce a bug in the browser.
Preconfigured for: On-call and bug-fixing sessions.
.mcp.json{
"mcpServers": {
"chrome-devtools": {
"command": "npx",
"args": [
"-y",
"chrome-devtools-mcp@latest",
"--no-usage-statistics"
]
},
"sentry": {
"type": "http",
"url": "https://mcp.sentry.dev/mcp"
},
"posthog": {
"type": "http",
"url": "https://mcp.posthog.com/mcp"
}
}
}
The hosted Stripe server (OAuth, so no API key in the file), Context7 for SDK docs and fetch for API reference pages, for building checkout and billing in USD.
Preconfigured for: SaaS and e-commerce teams adding checkout or subscriptions.
.mcp.json{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"stripe": {
"type": "http",
"url": "https://mcp.stripe.com"
},
"fetch": {
"command": "uvx",
"args": [
"mcp-server-fetch"
]
}
}
}
Vercel deployments and logs, Netlify site management and Cloudflare documentation, for shipping and debugging deploys.
Preconfigured for: Front-end and Jamstack teams that deploy to Vercel, Netlify or Cloudflare.
.mcp.json{
"mcpServers": {
"cloudflare-docs": {
"type": "http",
"url": "https://docs.mcp.cloudflare.com/mcp"
},
"vercel": {
"type": "http",
"url": "https://mcp.vercel.com"
},
"netlify": {
"command": "npx",
"args": [
"-y",
"@netlify/mcp"
]
}
}
}
Kubernetes in non-destructive mode (no deletes), AWS documentation and local git, for reading cluster state and preparing changes safely.
Preconfigured for: Platform and SRE engineers.
.mcp.json{
"mcpServers": {
"aws-docs": {
"command": "uvx",
"args": [
"awslabs.aws-documentation-mcp-server@latest"
],
"env": {
"FASTMCP_LOG_LEVEL": "ERROR"
}
},
"git": {
"command": "uvx",
"args": [
"mcp-server-git",
"--repository",
"."
]
},
"kubernetes": {
"command": "npx",
"args": [
"mcp-server-kubernetes"
],
"env": {
"ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS": "true"
}
}
}
}
Semgrep static analysis, GitHub for pull requests and code search, and fetch for advisories and CVE pages.
Preconfigured for: AppSec reviews and pre-merge security checks.
.mcp.json{
"mcpServers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/",
"headers": {
"Authorization": "Bearer ${GITHUB_PAT}"
}
},
"semgrep": {
"command": "uvx",
"args": [
"semgrep-mcp"
]
},
"fetch": {
"command": "uvx",
"args": [
"mcp-server-fetch"
]
}
}
}
Fetch pages as Markdown, search the web with Exa, keep notes in a local memory graph and reason step by step.
Preconfigured for: Research, competitive analysis and long investigations.
.mcp.json{
"mcpServers": {
"exa": {
"type": "http",
"url": "https://mcp.exa.ai/mcp"
},
"memory": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
},
"sequential-thinking": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-sequential-thinking"
]
},
"fetch": {
"command": "uvx",
"args": [
"mcp-server-fetch"
]
}
}
}
Only local servers and no accounts: a read-only filesystem, git for the current repository and a memory graph.
Preconfigured for: Offline or locked-down machines; a safe first setup.
.mcp.json{
"mcpServers": {
"git": {
"command": "uvx",
"args": [
"mcp-server-git",
"--repository",
"."
]
},
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"--read-only",
"."
]
},
"memory": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
}
}
}
Hugging Face for models, datasets and Spaces, Context7 for library docs, and fetch for papers and model cards.
Preconfigured for: ML engineers choosing and integrating models.
.mcp.json{
"mcpServers": {
"context7": {
"type": "http",
"url": "https://mcp.context7.com/mcp"
},
"huggingface": {
"type": "http",
"url": "https://huggingface.co/mcp"
},
"fetch": {
"command": "uvx",
"args": [
"mcp-server-fetch"
]
}
}
}
Neon serverless Postgres, postgres-mcp for query plans and index advice in restricted mode, and PostHog for product data.
Preconfigured for: Teams running Postgres-backed products with product analytics.
.mcp.json{
"mcpServers": {
"postgres": {
"command": "uvx",
"args": [
"postgres-mcp",
"--access-mode=restricted"
],
"env": {
"DATABASE_URI": "${DATABASE_URI}"
}
},
"neon": {
"type": "http",
"url": "https://mcp.neon.tech/mcp"
},
"posthog": {
"type": "http",
"url": "https://mcp.posthog.com/mcp"
}
}
}
The Model Context Protocol (MCP) is the open standard coding agents use to call outside tools: a browser, a database, your issue tracker, current documentation. Each tool runs as an MCP server, and each agent reads a config file that lists which servers to start or connect to. A local server is a command the client launches, such as npx @playwright/mcp@latest. A remote server is a URL the client connects to over HTTP, usually with an OAuth sign-in.
The catch is that every client wrote its own file format. They agree on the idea and differ on the details: the top-level key, whether a remote entry needs a type, how an environment variable is written, and where secrets go. Copying a snippet from a server's README into the wrong client is the most common reason a server silently fails to load. This generator keeps one list of servers and writes each client's file in that client's own syntax.
| Client | File | Shape | Secrets |
|---|---|---|---|
| Claude Code | .mcp.json | mcpServers; remote entries need "type": "http", or Claude Code treats them as local commands and skips them | ${VAR} or ${VAR:-default} in env, url and headers |
| Cursor | .cursor/mcp.json | mcpServers; local entries say "type": "stdio", remote entries are a url | ${env:VAR} |
| VS Code | .vscode/mcp.json | servers plus an inputs list; every entry has a type | ${input:id}: VS Code prompts once and stores the value securely |
| Codex | .codex/config.toml | TOML tables named [mcp_servers.name] | env_vars forwards variables; bearer_token_env_var for HTTP tokens |
Sources, read 2026-11-14: Claude Code MCP, Cursor MCP, VS Code MCP servers and Codex MCP. Codex reads a project .codex/config.toml only for trusted projects; otherwise merge the tables into ~/.codex/config.toml.
An MCP config is usually committed, so it must never contain a key. The generator writes only references, in each client's own syntax:
"BRAVE_API_KEY": "${BRAVE_API_KEY}" for Claude Code, ${env:BRAVE_API_KEY} for Cursor, an env_vars line for Codex.GITHUB_PAT. Give that token the smallest scopes the work needs.The test suite checks every template for strings that look like real keys (GitHub, Stripe, OpenAI, Slack and AWS formats) and fails the build if one appears.
37 servers in seven groups: 20 local servers started with npx or uvx, and 17 remote servers. Every package name was checked against the npm or PyPI registry on 2026-11-14, and every remote URL was probed and answered as a live MCP endpoint. Command-line flags come from each package's own README.
Packages that their maintainers have deprecated or archived are left out on purpose, even though older tutorials still recommend them: @modelcontextprotocol/server-postgres, @modelcontextprotocol/server-github, @modelcontextprotocol/server-brave-search, @stripe/mcp, @neondatabase/mcp-server-neon. Each has a maintained replacement here: postgres-mcp, the GitHub and Stripe remote servers, the new Brave Search package and Neon's remote server.
uvx. The rest run with npx and need Node.js.--read-only to the filesystem and local Supabase servers, --readOnly to MongoDB, --access-mode=restricted to postgres-mcp, and ALLOW_ONLY_NON_DESTRUCTIVE_TOOLS to Kubernetes.--no-usage-statistics, so it does not send usage data.Fewer than you think. Every server adds its tool descriptions to the agent's context on every task, and every server can act with your permissions. A focused set of three to six servers usually beats a long list.
Yes. The form, the four files and the template downloads are built in this page. Nothing you type is uploaded, stored or added to the address bar, and the files are the same for the same choices every time. The page never connects to any of the servers it lists.
For a project, .mcp.json at the repository root, with servers under mcpServers. Remote entries need "type": "http"; an entry with a url and no type is read as a local command and skipped.
Not reliably. The clients differ in the top-level key, in how remote entries are typed and in how environment variables are written. This generator writes each client its own file from one list.
In environment variables on your machine, or in VS Code's secure input storage. The generated files only reference them, so they are safe to commit.
The @modelcontextprotocol/server-github package is archived and unmaintained. GitHub runs the official server at api.githubcopilot.com/mcp/, which takes a personal access token.
It adds each server's own read-only flag where one exists: filesystem and local Supabase, MongoDB, postgres-mcp restricted mode and non-destructive Kubernetes. Servers without such a flag are unchanged.
You do not have to take our word for it. Developer tools show every request a page makes, so you can watch this page work and confirm that nothing you type is sent. The check takes under a minute.
F12, or Cmd+Option+I on a Mac, and select the Network tab.Only the feedback form can add a line to that log, and it sends only the vote and comment you type into it. 2prompt has no server code at all: no API, no model and no database of your files. The production site also sends a Content-Security-Policy header that tells your browser to refuse connections to anywhere except this site, the fonts and the anonymous analytics endpoint.