Every MCP client keeps its server list in its own file: Claude Code uses ~/.claude.json and a project .mcp.json, Claude Desktop uses claude_desktop_config.json, Cursor uses ~/.cursor/mcp.json, VS Code uses .vscode/mcp.json, Gemini CLI uses ~/.gemini/settings.json, and Codex uses ~/.codex/config.toml. Most clients also have project and managed scopes, and connectors added in a vendor’s web app have no local file at all. This page is the reference table. Why shadow MCP servers are hard to track is the hub.
Paths move. Last checked November 15, 2026. Claude Code user, project, and managed paths were re-checked against Anthropic’s docs on that date. The other rows match the vendor docs linked under the table.
Where does each client keep the file?
| Client | Scope | Path | Format key |
|---|---|---|---|
| Claude Code | user / local | ~/.claude.json | mcpServers |
| Claude Code | project | .mcp.json at the repo root (often committed) | mcpServers |
| Claude Code | managed | macOS: /Library/Application Support/ClaudeCode/managed-mcp.json; Linux and WSL: /etc/claude-code/managed-mcp.json; Windows: C:\Program Files\ClaudeCode\managed-mcp.json | allowedMcpServers / deniedMcpServers |
| Claude Desktop | user | macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; Windows: %APPDATA%\Claude\claude_desktop_config.json | mcpServers |
| Cursor | global | ~/.cursor/mcp.json | mcpServers |
| Cursor | project | .cursor/mcp.json | mcpServers |
| VS Code | workspace | .vscode/mcp.json | servers |
| VS Code | workspace | .mcp.json | mcpServers |
| VS Code | user profile | profile mcp.json (one per profile) | mcpServers |
| VS Code | dev container | customizations.vscode.mcp in devcontainer.json | MCP config block |
| VS Code / Copilot | Agent Host | ~/.copilot/mcp-config.json | Copilot MCP config |
| Windsurf (Devin Cascade) | user | macOS and Linux: ~/.config/devin/mcp_config.json; Windows: %APPDATA%\devin\mcp_config.json | MCP config |
| Gemini CLI | user | ~/.gemini/settings.json | mcpServers |
| Gemini CLI | project | .gemini/settings.json | mcpServers |
| OpenAI Codex | user | ~/.codex/config.toml | [mcp_servers.<name>] |
| Claude.ai custom connectors | account | no local file (connection originates from Anthropic’s cloud) | n/a |
| Cursor team / marketplace servers | team | no local file unless someone also wrote them into mcp.json | n/a |
Docs: Claude Code, managed MCP, Claude Desktop, Cursor, VS Code, Windsurf / Devin, Gemini CLI, Codex, Claude connectors.
Where does Claude Code store MCP servers?
User and local scope
User and local servers live in ~/.claude.json. Local scope is the default for claude mcp add. It is stored under that project’s path inside ~/.claude.json, so a server you add in one repo does not automatically appear in the others. User scope, passed with --scope user, is also in ~/.claude.json and is available across projects.
Project scope
Project servers live in a .mcp.json file at the repo root, under an mcpServers key. Teams often commit this file so everyone gets the same tools. Claude Code prompts for approval in an interactive session before using those servers.
Managed scope
Managed allowlists and denylists, and the exclusive managed server set, live in a system path that differs by operating system. The current filename is managed-mcp.json on all three platforms, in the directories in the table. allowedMcpServers and deniedMcpServers key on serverUrl or serverCommand. A serverName entry is not a security control.
Where is the Claude Desktop config?
On macOS it is ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows it is %APPDATA%\Claude\claude_desktop_config.json. Local servers in that file use the machine’s local network. Remote connectors you add in the Claude app are not stored there.
Where is Cursor’s mcp.json?
Global servers are in ~/.cursor/mcp.json. Project servers are in .cursor/mcp.json inside the repo. Both use the mcpServers key. Cursor’s docs also describe installing servers from the Marketplace or from a team marketplace, and those team or marketplace servers do not appear in either file unless someone also adds them there.
Where do VS Code and GitHub Copilot keep MCP config?
The workspace file .vscode/mcp.json uses a servers key, not mcpServers. A workspace .mcp.json uses mcpServers. User-profile config lives in a per-profile mcp.json. Dev containers can also carry a customizations.vscode.mcp block in devcontainer.json. Copilot’s Agent Host has its own file at ~/.copilot/mcp-config.json. VS Code can also read config that other tools wrote, so a sweep of only one of these paths will miss the rest.
Where did Windsurf’s config go?
If you install Windsurf, the MCP config path says Devin, because Windsurf’s docs now live under Devin and name ~/.config/devin/mcp_config.json on macOS and Linux, or %APPDATA%\devin\mcp_config.json on Windows. A sweep script written against an older Windsurf path will skip it. That is a small, true oddity, and exactly the kind of thing that breaks a scanner you wrote last spring.
Where does Gemini CLI store servers?
User servers go in ~/.gemini/settings.json. Project servers go in .gemini/settings.json. Both use mcpServers.
Where does OpenAI Codex keep MCP servers?
In ~/.codex/config.toml, as one [mcp_servers.name] table per server. That is TOML, not JSON, so a sweep that only parses JSON will miss it.
Which servers have no local file?
Claude custom connectors, and Cursor team or marketplace servers that were never also written into mcp.json, have no file on the laptop. The connection comes from the vendor’s cloud or from the team account. A file sweep will report a clean machine and still miss the server the developer uses every day.
How should you read these files safely?
Parse JSON for mcpServers, servers, and projects.*.mcpServers, and parse TOML for mcp_servers. Emit the client, the scope, the file, the server name, the transport, and the URL host or command basename. Never print env or headers values. Keys don’t belong in these files.
Where does one remote server land in three clients?
I use Pocket’s MCP server as the example, because I run it from Claude Code. The URL is https://public.heypocketai.com/mcp. In Claude Code a user-scope entry lands in ~/.claude.json under mcpServers. In Claude Desktop it lands in claude_desktop_config.json under the same key. In Cursor it lands in ~/.cursor/mcp.json, or in .cursor/mcp.json if I added it only for one repo. The meeting recorder I use is the product on the other end of that URL. If I add the same server as a Claude custom connector instead, none of those files will show it.
claude mcp add writes to local scope by default, which is the current project’s entry inside ~/.claude.json, unless you pass --scope project or --scope user.
What has changed in these paths?
- September 28, 2026: first table published on the shadow MCP hub, including Windsurf’s docs moving under Devin and naming
~/.config/devin/mcp_config.json. - November 15, 2026: Claude Code managed config is
managed-mcp.jsonon macOS, Linux, WSL, and Windows. Earlier shorthand on the hub listed the Linux and Windows directories without the filename. Anthropic’s managed MCP docs are the source for the filename.