Reusing one MCP setup across Claude, VS Code and Cursor

You can reuse MCP servers across clients in three documented ways: Claude Code can import from Claude Desktop, VS Code can discover configs from other apps, and a shared root config file is read by more than one tool. Each has limits, so check what was imported before you trust it.

Import into Claude Code

Claude Code has a command that imports servers from Claude Desktop. The docs note it is available on macOS and WSL.

claude mcp add-from-claude-desktop

You can also run Claude Code itself as an MCP server with the serve command, which lets another client call it.

Discovery in VS Code

VS Code has a setting that turns on discovery of configs from Claude Desktop, GitHub Copilot CLI, Cursor and Windsurf. The GitHub Copilot guide shows the same setting for reusing a Claude Desktop config.

Discovery means a server defined for one app can start appearing in another. That is convenient, and also a reason to audit the combined list. Trust dialogs still apply to servers from other sources.

A shared root file

VS Code documents a file named .mcp.json at the workspace root with a top-level mcpServers key, described as shareable. It notes that Copilot harness sessions read it directly. Claude Code uses the same file name for project scope. Cursor uses its own project file under a dot cursor folder, so it does not share this one.

The shapes are close but not identical. VS Code's own file uses a servers key, Claude and Cursor use mcpServers, and variable syntax differs. Claude Code expands dollar brace variables, Cursor uses an env prefix form. A server that works in one may need small edits in another.

Things that do not carry over

• Claude's web connectors are configured in your account, not in a file, and they connect from Anthropic's cloud. A local stdio server in a desktop config file is a separate mechanism and is not available on claude.ai or in Cowork.
• Approval and trust state is stored per client. Approving a server in one does not approve it in another.
• Secrets. Reference variables per machine rather than copying values.

A tidy approach for a team

1. Pick one shared project file for servers everyone needs, with no secrets in it.
2. Document the extra edits each client needs in the README.
3. Turn discovery on only if you review the merged server list.
4. List servers per client from time to time, and remove ones nobody uses.

For why the list matters, see inventory of MCP servers. Safeguard's asset discovery and MCP Guardrail features address this on the organisation side.

Quick check

Sources

• Claude Code MCP documentation
• VS Code MCP servers
• GitHub Copilot MCP setup
• Cursor MCP documentation
• Claude custom connectors with remote MCP

Last reviewed: 19 October 2026.