webtomcp
← docs

// docs / connect-chatgpt

Connect WebToMCP to ChatGPT in 2 minutes

This tutorial gives ChatGPT live access to any website you've indexed with WebToMCP. As of May 2026, custom MCP connectors work on every ChatGPT tier — Free, Plus, Pro, Business, Enterprise — so you can do this without paying OpenAI a cent. End-to-end time: about two minutes.

What you'll need

Heads up — read vs write actions: Custom MCP connectors are read-only on every tier. They can answer questions about your indexed site, but they cannot change anything on it. (Writing to external systems through ChatGPT actions is a separate workflow that needs Business/Enterprise.)

Step 1 — Create a knowledge base

  1. Sign in at app.webtomcp.net.
  2. Click + New knowledge base, paste your URL, and click Add.
  3. Wait 1-10 minutes for indexing. The dashboard polls live and shows page count climb. Status flips to ready when done.

Tip — path-prefix scoping: if you submit https://example.com/help/ we only crawl URLs under /help/. We won't bleed into /blog or /pricing on the same host. Keeps each knowledge base focused and predictable.

Step 2 — Copy your MCP URL

From the knowledge base detail page click Copy MCP URL. The URL looks like:

https://mcp.webtomcp.net/batches/<batch-id>/mcp?key=wmcp_<url-key>

Both the batch id and the URL key are scoped to your tenant — anyone with this full URL can query the endpoint, so treat it like a personal access token. Don't commit it to a public repo.

Step 3 — Add the connector inside ChatGPT

The fastest path is the CLI:

npm i -g webtomcp-cli
wmcp connect chatgpt

The CLI opens chatgpt.com in your default browser with the connector form pre-populated — URL, name, description all filled in. Click Connect in the ChatGPT dialog and you're done.

Doing it by hand

  1. In ChatGPT, open the conversation composer and click the + (or paperclip) icon.
  2. Choose Add connector (called Custom connectors on Free / Plus, and Connectors on Pro / Business / Enterprise).
  3. Pick MCP server from the connector type list.
  4. Paste the URL you copied in Step 2 into the Server URL field.
  5. Give the connector a recognisable name — e.g. my-docs, shopify-store, notion-wiki.
  6. Add an optional description so future-you remembers what this connector indexes.
  7. Click Save. The connector appears in the model-picker chip area.

Step 4 — Verify it's working

Start a new chat. Click the connector chip next to the message box and toggle my-docs on. Then ask something that only the indexed site would know:

"What's the refund policy on this site?" — or — "Summarize the API authentication flow."

If ChatGPT returns an answer with source URLs from the indexed pages, you're done. If you don't see the connector chip, refresh the chat once.

Troubleshooting

ChatGPT says "Connector not authorized"

The URL must include the ?key=wmcp_... query string. Re-copy from the WebToMCP dashboard — the bare batch URL without the key won't authenticate.

"Connector failed to start" / 500 error

Usually means the knowledge base is still indexing or hit an error during crawl. Check the dashboard — the batch should show ready. If it shows error, click into it to see the inline reason (most common: target site blocks crawlers, returns 403, or requires JS rendering for content that's static-fetched).

Connector chip doesn't appear in the chat composer

Refresh the chat tab once. If still missing, open ChatGPT in a new tab — the connector list is sometimes cached at session start. On Free tier, custom connectors only became GA in May 2026; if your client is older, update via cmd-shift-R.

Answers don't include source URLs

The system prompt may need a hint. Ask explicitly: "Use the connector and include source URLs in your answer." This usually nudges ChatGPT to attach citations.

Connector "times out" on every query

Could be a region issue — WebToMCP's MCP endpoints are served from Cloudflare's global edge, but ChatGPT routes through OpenAI's regions. Latency over 30s usually means a network mid-hop. Email developer@webtomcp.net with your batch id and we'll check the edge logs.

Answers feel stale — site updated recently

WebToMCP re-crawls weekly by default. Click Re-crawl in the dashboard to force an immediate refresh. The next ChatGPT query will hit the fresh snapshot.

Security & data custody

Indexed content is stored on Cloudflare (Workers, D1, Vectorize), encrypted at rest, never shared with advertisers, never used to train models. The MCP URL acts as a bearer token — anyone with the full URL (batch id + key) can query the endpoint, so guard it as you would any API key. If a key is compromised, regenerate it in the dashboard and update the ChatGPT connector. Full details at /legal/privacy.

What's next

Stuck? Email developer@webtomcp.net — we reply fast in the beta.