Point your own assistant at a real one
Not a video of our screen. A live MCP server you can connect to in about a minute and interrogate yourself. Ten governed tools over a fictional distributor's orders, receivables and stock, including the approval gate that guards every write.
One URL, no signup
This is the whole setup. There is no step two.
https://proceptio.com/api/demo/mcpAdd that as a custom connector in whichever assistant your team uses. No account, no email, no API key. It speaks the standard streamable-HTTP MCP transport, so any compliant client can reach it.
From the command line with Claude Code, it is one line:
claude mcp add --transport http proceptio-demo https://proceptio.com/api/demo/mcpOpen your assistant's connector settings
In Claude, Settings then Connectors, then add a custom connector. Copilot and ChatGPT have the equivalent under their own connector or integration settings.
Paste the URL above
Give it any name you like. The server tells your assistant what it can do, so there is nothing else to configure.
Ask it something a person would ask
Not "list your tools". Ask "who owes us the most and how old is it", the way somebody would on a Monday morning.
Read-only. Rate limited per visitor. Every figure belongs to a company that does not exist.
Ask it the Monday questions. That is what it is for.
Vantage Supply is an invented European distributor with six customers, seven open orders, a receivables ledger and a warehouse. Nothing here is anyone's real data.
“Who owes us the most, and how old is it?”
Watch it pick get_top_debtors on its own and come back
with a 30/60/90 ageing breakdown rather than a number with no provenance.
“Anything about to run out of stock?”
get_stock_cover returns weeks of cover against real
usage, and flags which shortages have no replenishment booked.
“Brief me on Meridian Steel before my call.”
One call to get_customer_briefing returns profile, open
orders, ageing and credit headroom. This is the meeting-prep shape our clients use most.
“Draft a payment reminder for them.”
The interesting one. It writes the message, then stops and waits. Keep reading below.
The part worth trying twice
Ask it to send the reminder. It will refuse, on purpose.
draft_payment_reminder is the only write-shaped tool on
the server, and it cannot write. Called normally it performs a dry run: it returns the
exact message that would go out, to which address, for which amount, plus a one-time
confirmation token. Nothing has happened yet.
Hand the token back and it reports what would happen next, including the audit entry that would be written and who it would name. Alter a single character of the token and it refuses, because an approval must not be replayable against different content.
That gate is the answer to the question every finance and compliance team asks first, and it is why the real deployments were allowed anywhere near live systems.
Dry run
Full message returned, nothing sent, token issued. A person can read exactly what the assistant proposes before anything is real.
Confirmed
Token matches, so the action is approved. In production this is where the send happens and the audit row is written. In this demo, nothing is sent or stored.
Tampered
Change the token and it is refused rather than guessed at. Try it: an approval that can be replayed is not an approval.
What this demo is not. Said plainly, before you ask.
Is the data real?
No. Every figure belongs to Vantage Supply, a fictional European distributor invented for this demo. No client data, and no anonymised client data, appears in it. Each tool result carries that disclaimer inside the payload, so your assistant repeats it when it summarises.
Can it change anything?
No. Every tool is read-only, and the one write-shaped tool cannot send or store anything under any circumstances. There is also no tool that runs arbitrary queries, here or in a real deployment. Named actions only, which is the whole architectural point.
Is this what we would get?
The shape, yes. The tools mirror what a real connector exposes and the approval gate is the same pattern. The difference is scale and specificity: our production servers carry between 15 and about 97 tools each, written against a particular company's systems and rules, reaching live data behind that company's own permissions.
Why is it free and ungated?
Because an email wall on a demo nobody has heard of collects addresses from the few people who already trust us and blocks everyone else. If it is useful, you know where to find us.
That took a minute and it was invented data. The same thing over your own systems is the actual conversation.
Compliance admin per quarter, across 7 funds and 191+ borrowers, on a 67-tool server built the way the one you just used was.
Tell us the one question your team asks every week and currently answers by opening three systems. In half an hour we can say whether an assistant would help with it, and whether the data behind it is in a fit state to be asked.
