The MCP config file for your client, written for you

Pick a server and a client and copy the exact command or file, in the shape that client reads, with the path it lives at. Or paste the file you have and read what is wrong with it, in words. Every shape comes from the client's own documentation, dated, and nothing you type leaves the page.

Server

Command or address as its documentation gives it. Edit anything below.

Name in the file

Kind

Address

Headers, one per line as Name=value

Client

One command in the terminal, or .mcp.json at the project root.

Terminal

$ claude mcp add --transport http --scope user sumibako https://sumibako.com/api/mcp

.mcp.json, at the project root, if you would rather commit it

{
  "mcpServers": {
    "sumibako": {
      "type": "http",
      "url": "https://sumibako.com/api/mcp"
    }
  }
}

The command writes to local scope unless you add --scope user or --scope project. In JSON, a url entry needs "type": "http"; without it Claude Code reads the entry as a local command and skips it. Run /mcp inside Claude Code to authenticate a remote server that asks you to sign in.

User, every project: ~/.claude.json, written by the command with --scope user
Project, shared with the repository: .mcp.json at the project root, or --scope project

Shape read from Claude Code's own documentation on 2026-10-30. If it has changed since, that page wins.

How it goes

01
Pick the server
Six are filled in from their makers' own documentation: a remote address or a local command with its arguments. Or describe your own: a name, an address for a remote server, or the command that starts a local one, with its arguments and environment variables.
02
Pick the client
Claude Code, Claude Desktop, Cursor, VS Code, Zed, JetBrains AI Assistant, Kiro, Codex or Gemini CLI. Each reads a different file in a different shape, and the output changes with the pick: the terminal command where the client has one, the file where it does not, and the path the file lives at.
03
Copy it where it says
The label on each block is the place: a terminal, a file in the project, a file in your home folder, a dialog in the settings. Under it is what that client does differently, said once, and the page the shape was read from with the date.
04
Or check the file you already have
Paste it, pick the client, and read what is wrong in words: the wrong top-level key for that client, an entry with neither a command nor an address, a url Claude Code would skip for want of a type, a token sitting in plain text in a file that is easy to commit.

What it knows

The top-level key each client reads: mcpServers for most, servers for VS Code, context_servers for Zed, a TOML table for Codex
Where the file lives in each client, project and user level, and the command for the two clients that have one
That Claude Code and VS Code want a type on every entry, and that Claude Code skips a url without one
That Gemini CLI reads httpUrl for an HTTP server and url for an SSE one, the one place the two keys differ
Kiro's disabled and autoApprove keys, Cursor's ${env:NAME} references, Codex's env table

What it leaves out

Windsurf. Its documentation moved on the day this was written and the page would not render, so its shape is not here rather than remembered.
ChatGPT and Claude.ai, which take a remote server as a connector in their settings, not as a file. The Claude guide and the ChatGPT guide on this site have the clicks.
OAuth client ids and secrets, input variables and sandboxes. Each client documents its own, and the link under the output goes there.
Checking TOML. The checker reads the JSON files; the Codex table is written on the other tab.

Questions

Is anything I paste sent anywhere?: No. The generator and the checker are a few hundred lines of code that run in this tab, there is no request to any server from this page, and the one thing recorded is which client was picked, so the next version knows which to keep current. A file with a token in it stays in your browser; the checker warns you that it is there.
Why does Claude Code refuse my url entry?: Because it has no type. Claude Code reads an entry without a type as a local command, and an entry with a url and no command is nothing it can start, so it skips the server and says so. Add "type": "http" (or "sse" for an SSE endpoint). The command line form, claude mcp add --transport http, sets it for you.
mcpServers or servers?: mcpServers for Cursor, Kiro, Gemini CLI, Claude Desktop and Claude Code's JSON. servers for VS Code's own mcp.json, and VS Code's reference says not to copy a file between the two formats without checking it. Zed uses context_servers in its settings file, and Codex a [mcp_servers.name] table in TOML. Paste the file into the checker with the client picked and it names the key.
Where is claude_desktop_config.json?: On macOS at ~/Library/Application Support/Claude/claude_desktop_config.json, on Windows at %APPDATA%\Claude\claude_desktop_config.json, or from the app through Settings, Developer, Edit Config. Its documented entries start a local command, and paths in the arguments have to be absolute. A remote server is added to the app as a custom connector in Settings instead.
How current are the formats?: Each client's shape was read from its own documentation on the date shown under the output, 2026-10-30 at the latest, and that page is linked so you can see for yourself. The formats change: a type became required in Claude Code, VS Code added a portable file beside its own, Windsurf's documentation moved. If a client's page has moved on, its page wins, and the hello address in the footer is where to say so.
Which server should I add first?: The one your work needs; the six presets are examples from the vendors' own pages, not a ranking. Sumibako's is there because it is ours: a remote server that lets an agent file pages into a vault and hand back a link, with no token to paste, and the agents guide on this site is where that is explained.

What an MCP connector can read and write once it is in, and what to check before adding one, is in What an MCP connector can do to your data. Adding a server to Claude.ai or ChatGPT, which take a connector rather than a file, is in the Claude guide and the ChatGPT guide.

One server worth the entry

Sumibako's server is the first preset: an agent that can file its plan, its handoff or its research as a page in your vault and hand you a link, from any of the clients above, with no token to paste. The free plan is enough to find out.

Start writing - free How the agent connection works

How it goes

Pick the server

Six are filled in from their makers' own documentation: a remote address or a local command with its arguments. Or describe your own: a name, an address for a remote server, or the command that starts a local one, with its arguments and environment variables.

Pick the client

Claude Code, Claude Desktop, Cursor, VS Code, Zed, JetBrains AI Assistant, Kiro, Codex or Gemini CLI. Each reads a different file in a different shape, and the output changes with the pick: the terminal command where the client has one, the file where it does not, and the path the file lives at.

Copy it where it says

The label on each block is the place: a terminal, a file in the project, a file in your home folder, a dialog in the settings. Under it is what that client does differently, said once, and the page the shape was read from with the date.

Or check the file you already have

Paste it, pick the client, and read what is wrong in words: the wrong top-level key for that client, an entry with neither a command nor an address, a url Claude Code would skip for want of a type, a token sitting in plain text in a file that is easy to commit.

What it knows

The top-level key each client reads: mcpServers for most, servers for VS Code, context_servers for Zed, a TOML table for Codex

Where the file lives in each client, project and user level, and the command for the two clients that have one

That Claude Code and VS Code want a type on every entry, and that Claude Code skips a url without one

That Gemini CLI reads httpUrl for an HTTP server and url for an SSE one, the one place the two keys differ

Kiro's disabled and autoApprove keys, Cursor's ${env:NAME} references, Codex's env table

What it leaves out

Windsurf. Its documentation moved on the day this was written and the page would not render, so its shape is not here rather than remembered.

ChatGPT and Claude.ai, which take a remote server as a connector in their settings, not as a file. The Claude guide and the ChatGPT guide on this site have the clicks.

OAuth client ids and secrets, input variables and sandboxes. Each client documents its own, and the link under the output goes there.

Checking TOML. The checker reads the JSON files; the Codex table is written on the other tab.

Questions

Is anything I paste sent anywhere?

No. The generator and the checker are a few hundred lines of code that run in this tab, there is no request to any server from this page, and the one thing recorded is which client was picked, so the next version knows which to keep current. A file with a token in it stays in your browser; the checker warns you that it is there.

Why does Claude Code refuse my url entry?

Because it has no type. Claude Code reads an entry without a type as a local command, and an entry with a url and no command is nothing it can start, so it skips the server and says so. Add "type": "http" (or "sse" for an SSE endpoint). The command line form, claude mcp add --transport http, sets it for you.

mcpServers or servers?

mcpServers for Cursor, Kiro, Gemini CLI, Claude Desktop and Claude Code's JSON. servers for VS Code's own mcp.json, and VS Code's reference says not to copy a file between the two formats without checking it. Zed uses context_servers in its settings file, and Codex a [mcp_servers.name] table in TOML. Paste the file into the checker with the client picked and it names the key.

Where is claude_desktop_config.json?

On macOS at ~/Library/Application Support/Claude/claude_desktop_config.json, on Windows at %APPDATA%\Claude\claude_desktop_config.json, or from the app through Settings, Developer, Edit Config. Its documented entries start a local command, and paths in the arguments have to be absolute. A remote server is added to the app as a custom connector in Settings instead.

How current are the formats?

Each client's shape was read from its own documentation on the date shown under the output, 2026-10-30 at the latest, and that page is linked so you can see for yourself. The formats change: a type became required in Claude Code, VS Code added a portable file beside its own, Windsurf's documentation moved. If a client's page has moved on, its page wins, and the hello address in the footer is where to say so.

Which server should I add first?

The one your work needs; the six presets are examples from the vendors' own pages, not a ranking. Sumibako's is there because it is ours: a remote server that lets an agent file pages into a vault and hand back a link, with no token to paste, and the agents guide on this site is where that is explained.