MCP config generator: one server, the right setup for 5 coding agents
Type one MCP server's name and its start command or URL, and this page writes the setup for Claude Code, Codex, Cursor, VS Code and Claude Desktop: the command to run, or the JSON or TOML entry to paste. We read each format from that agent's own docs on November 7, 2026. Nothing you type leaves your browser.
For builders who add one MCP server to more than one agent: each agent stores the same server in a different shape. VS Code keys it under servers, Cursor and Claude Desktop under mcpServers, Codex in a TOML table, and Claude Code takes one command. The example below is the UserProof server, which answered its first request 4.7 seconds after a cold start in our test.
Your server
Example result
5 agents for userproof, a local server started by npx -y github:GetUserProof/userproof.
Claude Code
Command run it in your project folder
claude mcp add --transport stdio userproof -- npx -y github:GetUserProof/userproofAdd --scope user for every project, or --scope project to write .mcp.json for your team.
.mcp.json in the project root, shared with your team
{
"mcpServers": {
"userproof": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"github:GetUserProof/userproof"
]
}
}
}Codex (CLI, IDE extension and ChatGPT desktop app)
Command
codex mcp add userproof -- npx -y github:GetUserProof/userproofOr ~/.codex/config.toml
[mcp_servers.userproof]
command = "npx"
args = ["-y", "github:GetUserProof/userproof"]
startup_timeout_sec = 60The default start timeout is 10 seconds; the first start of npx downloads the server and can take longer. The ChatGPT desktop app and the Codex IDE extension read this same file.
Cursor
~/.cursor/mcp.json (or .cursor/mcp.json in one project)
{
"mcpServers": {
"userproof": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"github:GetUserProof/userproof"
]
}
}
}Already have the file? Add the inner "userproof" entry under your existing mcpServers.
VS Code
.vscode/mcp.json in your workspace
{
"servers": {
"userproof": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"github:GetUserProof/userproof"
]
}
}
}VS Code's file uses servers, not mcpServers.
Claude Desktop
claude_desktop_config.json Settings, Developer, Edit Config (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\)
{
"mcpServers": {
"userproof": {
"command": "npx",
"args": [
"-y",
"github:GetUserProof/userproof"
]
}
}
}Quit and reopen Claude Desktop after saving.
Where each agent keeps its servers
| Agent | Where the server goes | Top-level key | Remote servers | Reads a variable as |
|---|---|---|---|---|
| Claude Code | claude mcp add; .mcp.json for a project | mcpServers | --transport http | ${NAME} |
| Codex, ChatGPT desktop app | codex mcp add; ~/.codex/config.toml | [mcp_servers.name] | url | env_vars list |
| Cursor | ~/.cursor/mcp.json or .cursor/mcp.json | mcpServers | url | ${env:NAME} |
| VS Code | .vscode/mcp.json | servers | "type": "http" | ${env:NAME} |
| Claude Desktop | claude_desktop_config.json | mcpServers | Settings, Connectors | value in the file |
Each row was read from the agent's own docs on November 7, 2026 (links under Sources). The ChatGPT desktop app, the Codex CLI and the Codex IDE extension share one file, so one entry covers all three.
How to use it
- Name the server. Letters, numbers,
-or_. Done looks like the name in every block below. - Say how it runs. A local server starts from a command such as
npx,uvxor a path, plus its arguments. A remote one is a URL. Done looks like the summary line naming the command or URL. - List the variable names it needs, if any. For a remote server, the one that holds your bearer token. Done looks like each agent's own way of reading it, such as
${env:API_KEY}for Cursor. - Copy the block for your agent. Paste the command in a terminal, or merge the entry into the file. Done looks like the server in the agent's list (
claude mcp list,codex mcp list, or/mcp).
Keep keys out of the files
The generator never asks for a value. Cursor and VS Code read ${env:NAME} from your environment, Claude Code's .mcp.json reads ${NAME}, and Codex's env_vars list forwards variables you already set. So the file you commit holds names only. Claude Desktop's guide writes values into its config file, so that block leaves a PASTE_..._HERE marker for you to fill in on your own computer. The UserProof server needs no variable at all: it reads its key from ~/.config/userproof/key, which setup writes.
Limits, stated plainly
- It writes the entry, not the server. It can't tell whether your command starts or your URL answers. Test the server first, then add it.
- Merge by hand. If the file already lists servers, paste the inner entry under the existing key instead of replacing the file.
- Two agents are left out. Gemini CLI's docs now say it was replaced by Antigravity CLI on June 18, 2026 for unpaid-tier and Google One users, and Windsurf's MCP docs link now opens Devin Desktop's docs, which don't show the file format. We add an agent once its docs show the format.
- Sign-in flows are the agent's job. A remote server that uses OAuth asks you to sign in when the agent first connects; nothing goes in the file for that.
What UserProof adds once it's connected
UserProof is the example server above. Once it is connected, it records what real visitors do on your site and hands your coding agent the one fix with the file and line, then checks whether it worked. Free covers 3 sites and 5,000 recorded visits a month; Pro is $29 a month (pricing, read November 7, 2026). Setup guides: Claude Code, Codex and Cursor, ChatGPT.
Related free tools: the dead click finder, the funnel calculator, the session replay cost calculator.
FAQ
What is the difference between mcp.json and .mcp.json?
They belong to different agents. Cursor reads mcp.json from ~/.cursor or a project's .cursor folder. Claude Code reads .mcp.json from the project root. VS Code reads .vscode/mcp.json and keys servers under "servers" instead of "mcpServers".
Where is the Codex MCP config file?
~/.codex/config.toml, or .codex/config.toml inside a trusted project. Each server is a [mcp_servers.name] table. The ChatGPT desktop app and the Codex IDE extension read the same file.
Why does my MCP server time out in Codex?
Codex gives a server 10 seconds to start by default. A server started with npx downloads itself the first time, which can take longer. Add startup_timeout_sec = 60 to its table; the generator does this for npx, uvx, bunx and pnpx.
Does this generator send or store what I type?
No. It runs in your browser, makes no requests and saves nothing. The page has the same visit analytics as the rest of this site, and the form and the result are marked so that analytics never records them. Reload the page and it starts from the example again.
Sources, read November 7, 2026
- Claude Code: Claude Code docs, MCP (
claude mcp add, scopes,.mcp.json,${VAR}expansion), read November 7, 2026 - Codex and the ChatGPT desktop app: OpenAI, Model Context Protocol (codex mcp add, config.toml keys, the 10-second default), read November 7, 2026
- Cursor: Cursor docs, MCP (
mcp.jsonlocations,${env:NAME}), read November 7, 2026 - VS Code: MCP configuration reference (
servers,type,${env:NAME}), read November 7, 2026 - Claude Desktop: Connect to local MCP servers and Connect to remote MCP servers, read November 7, 2026
- Left out: Gemini CLI MCP docs (replaced by Antigravity CLI notice) and Devin Desktop docs (where Windsurf's MCP link leads), read November 7, 2026
- The 4.7-second start: one cold start of the UserProof server (version 3.10.0) with an empty npm cache on our Mac mini, November 7, 2026. The generator's code is the open file mcp-config.js.
All of our free tools, including the dead click finder and the funnel calculator, are on the free tools page.