Add onelinktoken as a remote MCP connector and your assistant can mint, check, and kill one-time tokens on your behalf — in the conversation, with your own account, under an OAuth consent you grant yourself.
The connector is not live yet
This page documents the connector as it is being built. The MCP endpoint
([MCP endpoint URL]) is not serving traffic yet, so the steps below will
not work today. Nothing here describes a shipped feature — treat it as the spec for what connecting will
look like.
onelinktoken issues short-lived credentials: magic links, one-time passwords, and bearer tokens. The connector exposes those operations to an AI assistant as MCP tools, so instead of writing an API call you ask for what you want — "send a 15-minute login link to this address", "is that token still valid?", "kill it" — and the assistant calls the matching tool against your account.
It is a remote MCP server. There is nothing to install and no API key to paste into a config file: you add the endpoint in your assistant, sign in through the browser once, and the connector holds an OAuth grant scoped to your account.
The connector advertises exactly these seven tools. Your assistant discovers them automatically — you do not need to memorise the names, they are listed here so you know the full extent of what it can do.
Mints a single-use sign-in link for an address, with an expiry you choose.
Mints a short numeric one-time password for an email or SMS verification step.
Mints an API access token that can be revoked the moment you are done with it.
Spends a token once and reports whether it was valid. A redeemed token cannot be reused.
Kills a token immediately, whether or not it has been used.
Reports the state of one token — pending, redeemed, expired, or revoked.
Lists your recent tokens and their states so you can audit or clean up.
Claude supports remote MCP connectors. Availability of custom connectors depends on your Claude plan and, on a Team or Enterprise workspace, on whether an administrator has allowed them.
In Claude, go to Settings → Connectors and choose to add a custom connector.
Enter [MCP endpoint URL] as the server URL.
A browser window opens on onelinktoken. Sign in, read the consent screen, and approve.
The seven tools appear in the conversation. Ask for a token in plain language.
ChatGPT also speaks MCP for remote connectors. As with Claude, whether you can add your own connector depends on your plan and on any workspace policy your administrator has set.
In ChatGPT settings, find connectors and choose to add one by URL.
Enter [MCP endpoint URL].
The same onelinktoken sign-in and consent screen opens in a browser tab.
Turn the connector on for the conversation and the same seven tools are available.
The connector uses OAuth. You sign in to onelinktoken yourself — the assistant never sees your password — and you are then shown a consent screen that names the account and lists what the assistant is being allowed to do: create tokens, redeem them, revoke them, and read their status and history.
The grant is bound to the account you signed in with. The assistant can reach that account's tokens and nothing else — not another tenant's, not another user's.
Remove the connector in Claude or ChatGPT, or revoke the grant from your onelinktoken dashboard. Either ends the assistant's access.
Tool calls land in the same audit trail as REST calls — created, redeemed, revoked, with timestamps — so you can see exactly what the assistant did.
Read this before you mint anything sensitive
When the assistant calls create_magic_link,
create_otp, or create_bearer_token, the
resulting link, OTP, or token value is returned into the conversation. That
means it is written into your chat history, which the AI host — Anthropic or OpenAI, not us — stores on its own
servers under its own retention policy. It may also be visible to anyone the conversation is shared with, and to
anyone with access to that workspace's chat logs.
A live credential sitting in a stored transcript is a live credential. Treat it that way:
revoke_token when you are done, which makes the copy in the transcript useless.We store only a hash of every token, so we cannot show you the value again after creation — but that does nothing about the plaintext copy in the transcript. Deleting it or revoking the token is the only thing that does.
Connecting is free. A new account gets a monthly allowance of tokens that is meant for evaluating the service and for light personal use — enough to wire up a flow and see it work, not enough to run a product on. Paid plans raise that allowance substantially and add the features listed on the pricing page.
When you run out of free-tier tokens, the tool does not fail with a cryptic error. It returns a normal result that says the allowance is used up and includes an upgrade link. The assistant shows you the link; you decide whether to follow it. Nothing is charged and nothing is created until you do.
Current allowances and plan features are on the pricing section — that page is the single source of truth, so we don't restate the numbers here.
Payment never happens inside the chat. The upgrade link opens our own Stripe checkout page in a browser tab, where you enter your card details directly with Stripe as you would on any other website.
If a tool call fails, if the consent screen doesn't appear, or if you think a token has leaked, revoke first and ask questions second — then get in touch.
Support can pull the audit trail for a token ID and tell you exactly what the assistant did. Never send us the raw token value — the ID is enough.
Contact support →