Your AI agent can draft an invoice, pull a VAT number out of a PDF, and sound very sure that everything checks out.
That confidence is not a receipt.
If you build agents or MCP tools for EU B2B work, you already know the gap: the model can claim a VAT number is valid, or that a buyer is Peppol-reachable. What it often cannot show you is a dated trail of what was looked up, when, and with what result.
This post is for people shipping agents on Claude, ChatGPT, Cursor, Codex, Grok, and similar hosts. The agent stays the brain. What you wire in next is prepaid, read-only evidence.
Agents are great at language. Registries are not language.
VIES does not care how persuasive your prompt was. Peppol participant lookup is not a vibe check. Invoice readiness is not "looks fine in the chat."
When an agent says "VAT is fine," you want answers to questions like:
- Was the format check run, or did the model just recognize a familiar pattern?
- Was live VIES called, and on which date?
- Was the result valid, invalid, or unavailable (registry did not answer)?
- Did anyone look up Peppol reachability for that counterparty, or did the agent assume e-invoicing readiness from a website or a memory?
Without those answers, you have a fluent assistant. You do not have an audit trail.
Dated receipts matter because EU work is time-bound. Last month's VIES hit is not today's reverse-charge proof. "Probably on Peppol" is not a participant lookup with a timestamp.
What "receipt" means here
A useful tool receipt for an agent is something you can store next to the task:
- which tool ran
- which inputs were sent (normalized VAT / participant id, and so on)
- what the upstream check returned
- when it happened
- what it cost
That is "the agent said so" versus "the agent called a bounded tool and kept the proof."
Jithox: read-only prepaid MCP tools, not a new brain
Jithox does not replace Claude, ChatGPT, Cursor, Codex, or Grok. Those stay the reasoning layer.
Jithox is a set of read-only MCP tools billed in prepaid EUR, with clear limits, permissions, budgets, and receipts. You connect once to the MCP host, then your agent can call the tools when a workflow needs EU registry facts instead of guesses.
What is live today
| Tool | Price (accepted call) | What it is for |
|---|---|---|
| E-Invoice | €0.10 | VAT format checks, live VIES, Peppol participant lookup, invoice validation / readiness |
| Import | €0.25 | Import / customs preflight helpers |
| Energy Label | €0.25 | Energy label lookups |
| Sanctions | €0.50 | Sanctions screening lookups |
Failed or unavailable calls are €0. You are not charged for a registry outage dressed up as a success.
These tools are read-only. They do not execute payments, send invoices, or act as a Peppol Access Point. Your agent still decides what to do with the evidence. Jithox supplies the check and the receipt.
Why VIES + Peppol reachability belong in the same agent path
Many EU B2B flows need both:
- Is this VAT number structurally and registry-valid right now? Live VIES (plus format validation) answers that with a dated result, including unavailable when VIES cannot be reached.
- Can this party be reached on Peppol as a participant? Participant lookup is how you stop treating "they do e-invoicing somewhere" as "they are reachable for this network path today."
Agents that only do (1) still invent (2). Agents that only scrape a directory still invent (1). MCP tools with budgets and receipts keep the model on workflow language and the registries on facts.
Invoice validation / readiness sits beside those lookups so the agent can ask whether a document is shaped for the next step, without claiming anything was sent.
Connect once. Keep permissions and budgets boring.
MCP builders already fight Excess Agency: too many tools, too much write power, unclear spend. Keep it narrow:
- connect your host to
https://mcp.jithox.com/mcp(or start from connect your AI) - grant only the read-only tools you need
- set a prepaid budget so a runaway loop cannot surprise you
- require the agent to keep the tool receipt next to the business decision
For VAT and Peppol preflight, boring is a feature.
Try it without a card
There is a trial: 25 accepted calls / 14 days, no card. Enough to wire a Claude / Cursor / ChatGPT / Codex / Grok path, fire real VIES and Peppol participant lookups, and see what a dated receipt looks like in your own logs.
Failed and unavailable calls do not eat that budget the way a silent "success" hallucination would. That is deliberate.
Build-in-public note
We ship this for agent and MCP builders who need EU registry truth without replacing their model stack. No fake logos, no invented metrics. The offer is the tools, the prices, the receipts, and the trial.
If your agent still answers "VAT is fine" with only confidence and no dated trail, give it a tool that leaves a receipt.
Connect once: https://jithox.com/connect-your-ai
MCP host: https://mcp.jithox.com/mcp
Top comments (0)