How to Add MCP Servers in ChatGPT (2026)
ChatGPT connects to remote MCP servers as "custom connectors". It does not launch local stdio processes — if a server only runs locally, you need a tunnel or hosted deployment before ChatGPT can reach it.
1. Add a custom connector
In ChatGPT, open the Plugins page (chatgpt.com/plugins), click the plus button and choose "Add custom MCP server". A legacy flow also exists under Settings → Apps → Create.
Paste your server's MCP URL and choose its authentication. Depending on plan, creating connectors may be limited to admins/owners (Business) or admins plus RBAC-authorized developers (Enterprise/Edu).
2. Know the plan limits
Business and Enterprise/Edu get full MCP support on ChatGPT web (not mobile). Pro can build Apps SDK apps and connect custom MCP servers with read/fetch permissions only.
Deep Research can use connected apps for read/fetch actions only, not write actions. After publishing, ChatGPT uses a frozen snapshot of the tools until an admin refreshes it.
3. Local servers need a tunnel
Because ChatGPT only speaks to remote servers, a local stdio server must be exposed over HTTP. OpenAI's guidance is to use a secure tunnel (their "Secure MCP Tunnel" tooling) rather than exposing a dev machine directly.
What to serve
Your endpoint must speak MCP over HTTP. The earlier requirement that deep-research connectors expose exactly search and fetch tools has been relaxed — connected servers no longer need those specific tools, but read/fetch-only still applies to Deep Research.
{
"mcpServers": {
"my-connector": { "url": "https://your-host.example.com/mcp" }
}
}Skip the JSON
The MCPNav config generator builds a ready-to-paste config for ChatGPT from any servers you pick — including optional version pinning. Or start from our curated shortlists: ChatGPT connectors guide.