Add Replilo to Claude as a custom connector

Open Claude's connector settings, add a custom connector, paste https://replilo.com/mcp, and approve the page that appears. That is the whole setup. There is no key to generate, no client ID to register, no JSON to edit. Afterwards Claude can upload a document to your account, put it behind a verified-email wall and hand you back a share link, in the same conversation where it wrote the thing.

What the connector gives Claude

The same account you use in the dashboard, reached over the Model Context Protocol. Claude can list your documents, upload a new one, create and edit share links, set an expiry or a download cap, rotate a slug that leaked, and read your subscribers and download history. It is the dashboard's surface area, available to something that can act on it without you clicking through.

The reasoning behind that — why an assistant with a finished report has no way to hand it to anyone, and why large files never travel through the model — is in how MCP turns a report into a share link. This post is just the setup.

Adding it

1. Find the connector settings

In claude.ai, custom connectors live in settings under Connectors. The exact menu path and wording shift between surfaces and plans, so look for the option to add a custom connector rather than following a screenshot. If your plan doesn't offer custom connectors at all, skip to the API key route further down — it reaches the same server.

2. Paste the URL

https://replilo.com/mcp

That is the entire configuration. Claude fetches /.well-known/oauth-protected-resource/mcp, learns which authorization server to talk to, and registers itself as a client on the spot. Dynamic client registration is what removes the copy-paste step that normally comes with connecting two services: nobody has to provision a client ID in advance because the client creates its own.

3. Approve once

You are sent to Replilo to approve. If you're not already signed in you'll sign in first — the approval has to be made by the account owner, not by whoever happens to be holding the link. The page names the client asking, states which of your accounts it would act as, lists what it will be able to do, and prints the host it will return you to.

Click Connect and you land back in Claude with the connector live. Cancel and nothing is granted.

4. Ask it for something

"Upload this as a gated download and give me the link." Claude uploads the file, creates the link, and replies with a https://replilo.com/s/… URL you can paste anywhere. Anyone who opens it gives an email address and confirms it with a six-digit code before the file moves.

What you actually approved

One scope, and it is all-or-nothing. The consent page lists three lines because that is what the single scope covers — not because they are three separate switches you can grant individually. An approved client can read your subscriber list and delete your documents. If that is more than you want a particular client holding, don't approve it.

The one screen worth reading

The client's name on that page is whatever the client called itself when it registered. Anything can register, and anything can call itself anything — so a familiar name on its own proves nothing. The part that cannot be faked is the host you will be returned to, which is why the page prints it in plain text rather than tucking it behind the button.

The rule that follows is short: only continue if you started this from that client yourself, a moment ago. An approval page you arrive at from a link someone sent you is the same trick as any other phishing page, wearing an OAuth costume. Nothing legitimate needs you to approve a connector you didn't just ask for.

If your client can't use connectors

Custom connectors exist because browser-based clients cannot attach an Authorization header. Anything that can — Claude Code, a script, an agent framework — has a simpler path: create an API key on the AI agents page and send it as a bearer token.

Authorization: Bearer rpl_…

The key is shown once. Only a SHA-256 of it is stored, so it cannot be recovered afterwards — losing it means deleting it and making another, which is also how you revoke one. Keys are deliberately refused on the key-management endpoints themselves, so an agent holding a key cannot mint itself a second credential.

When it doesn't connect

Three things account for most failures. The plan you're on may not offer custom connectors, in which case no amount of retrying the URL helps — use an API key. You may be signed into the dashboard as a different account than the one you meant to grant, which the consent page tells you if you read the address it names. Or the grant was revoked, in which case the connector fails on its next refresh and needs approving again.

If you run Replilo yourself rather than on replilo.com, there is a fourth: your proxy has to route /oauth and /.well-known to the API. Discovery fails silently-ish otherwise, and the client reports only that it could not authenticate.

The short version

Paste https://replilo.com/mcp into Claude's custom connector settings and approve once. Check the host on the approval page and make sure you're the one who started it. What you grant is broad and revocable from the dashboard, and anything that can set a header should use an API key instead.