Local MCP configuration

mcp-sync

Define your MCP servers once. Keep every supported local AI client in sync.

mcp-sync gives local MCP configuration one versioned source of truth, previews every change, and preserves settings owned by each client.

Read the documentation View source on GitHub

The hidden problem

Your MCP server is configured once per AI app.

Claude Desktop, Cursor, Windsurf, VS Code, Codex, and Kiro use different files and native shapes. Manual copying creates drift, stale values, malformed configuration, and edits that are hard to review or undo.

How it helps

How it helps

  1. 01

    Define once

    Import compatible local definitions or add a complete local STDIO server to canonical versioned configuration.

  2. 02

    Preview the exact plan

    Validate and render every supported target in memory before the first write, with sensitive values structurally redacted.

  3. 03

    Apply recoverably

    Preserve client-owned fields, skip no-ops, retain adjacent backups, and roll back earlier writes when a later target fails.

What it covers

What it covers

  • Claude Desktop and Cursor
  • Windsurf and VS Code
  • Codex and Kiro
  • Dry-run and deterministic plans
  • Bounded named-server health checks
  • Backups, restore, locking, and rollback

Deliberate boundary

Configuration only—until you explicitly test.

init, list, sync, and restore do not start MCP servers. Only test executes one named local definition. Project-level files, remote entries, credential stores, and unowned client settings stay outside the managed boundary.