Full Webinar Transcript
Introduction and what we'll cover
We're going to go through what MCP is and why it should matter, a few example queries we can run in chat tools such as ChatGPT or Claude, and by the end of it we'll all be able to type a complex query such as "summarise my receipts for tax prep" and get a complete document ready to be sent to accountants, complete with deductions, categories and so on.
There's no technical background needed, and that's the whole point. It's going to be very simple to run. Some of the queries I'll show are actually inspired by our customers, custom dashboards I've been fascinated by, so a few of those will appear here. We'll end with a short Q&A if anyone wants to ask questions, otherwise we'll just close it once we've gone through all the parts.
What is MCP?
Let's start by saying what MCP is. MCP stands for Model Context Protocol. It's an open standard created by Anthropic a couple of years ago, and then everyone else, like OpenAI, adopted it. It became the plumbing the whole industry converged on.
If I were to compare it to something, I'd compare it to USB-C. Nowadays every device is compatible with USB-C: you plug it in, you don't need any adapter. Back in the day every device had a different connector, and that was a mess. Similarly, if you want to connect services and data sources to your favourite chat interface like Claude or ChatGPT, you use MCP. It's just a weird name for a way to standardise the connection between the chatbot and external data. You never actually see the protocol itself, so an easier way to refer to it is simply an AI connector.
Why connect Receiptor AI via MCP?
In the context of expense management, you might want to do this for at least two reasons.
The first is for your own personal dashboards, where you want to see where your expenses are going, see forecasts, and have custom views for yourself.
The second is to give your accountant or bookkeeper the report they're usually chasing you for. Receiptor already automates a lot of this. Previously you had to go through your emails and various sources, extract things manually, and type them into your accounting software. Receiptor already does that, but there was still some manual work left if you wanted to send a monthly report to your accountant. That can be automated further by connecting Receiptor to a chat interface like Claude or ChatGPT and asking in natural language.
The old way required you to log in, set filters in Receiptor, export to CSV or Excel, check the formatting, build the report, and email it to your accountant. The new way I'm about to show you lets you create a monthly expense report just by typing into the chat. I've got a few queries here that I'll type into Claude (or ChatGPT) so we can see the results live.
Live demo: top expenses
I'll send this query to Claude, and it will automatically try to find the tools required to answer it. I'll show you later how to connect Receiptor to Claude, but here it's just going through the tools it found and, in a few seconds, outputting a table with the result.
This is using Anthropic's model intelligence to connect to Receiptor AI data, which serves as a container of everything extracted from your emails, WhatsApp and other sources, exactly as you set it up. That's really the power of this.
As you can see, the top expenses here are our own database and server bills, plus some flights, an Intuit partnership and so on. The query can be adjusted very easily to get more documents, fewer documents, or to perform an action.
"What can Receiptor AI do?" and the system of record
Here's something I hadn't even planned to show: "Show me what Receiptor AI can do." Because Claude knows we have different tools serving different purposes, it fetches the server info and then the capabilities. It tells us we can perform actions around our documents, run exports and integrations, organise merchants, and create views in Receiptor.
I want to highlight this: what Receiptor is nowadays is an infrastructure, a system of record containing all the data and documents that have been extracted. When you log into Receiptor, that's just one of the interfaces. All of those key actions can also be performed through this connection via Claude or ChatGPT, and similarly via WhatsApp if you've enabled the AI feature in the mobile scanners section of our app. You can ask queries exactly as I'm doing now, directly from your mobile number on WhatsApp.
Custom dashboards from Receiptor data
Moving to the next example: "Find all invoices from Calendly this quarter." It'll run through that while I show you something that takes a little longer, building a custom dashboard out of Receiptor data, which I created earlier.
My first query was "Create a dashboard of my expenses by category and vendor." What's extraordinary is that even without me mentioning Receiptor, the model doesn't actually know in advance whether I want it to call Receiptor, but in the context of the tools I've made available it understands that the expenses relate to Receiptor. So it found the right tool and automatically built a live dashboard with Receiptor AI data that can be refreshed live, containing all the data living in your workspace.
This highlights one thing in particular: Receiptor holds the extracted data, the receipts and invoices, but just by connecting it to Claude via MCP, custom visualisations become possible. I can imagine this as a way for, say, a restaurant to filter documents by day and by category, so the stock-taker can have a custom view of what's going out, what's missing in stock, and what's coming in tomorrow, and try to see the forecast.
Another example is travel expenses by entity, by month and by merchant, which I've run before. By "entity" I mean this: if you have different business entities, the system identifies the differences between them and categorises them accordingly. In this dashboard it's grouping expenses by entity and by month, exactly as requested. I built a custom view of travel expenses by business entity, month and merchant, with a modifiable date range, and it produced this dashboard I can use myself.
One thing worth mentioning: I've connected another MCP that lets me push this website, which currently lives on my computer, to a hosted website, so I get a link I can share with others. For this demo I'll skip it, but it can be done through something like Vercel: I'd just say "push this website through Vercel" and get back a link to host and share it.
Headline use case: summarise receipts for tax prep
Now let's go straight to "Summarise my receipts for tax prep." I ran this earlier because it fetches all the data from Receiptor and takes a few minutes.
What it did was summarise my receipts for tax prep: it ran for a few minutes, created temporary local files within its environment, and produced a workbook covering all the documents from 2025 in my workspace, verifying the totals with deductible candidate expenses. The result is staggering. It created an Excel file I can save to Google Drive or export locally, with all the receipts and invoices, a tax prep summary, the total of all documents captured in 2025, all the categories and accounts, and all the transactions, even with links back to Receiptor. (This particular workspace was incomplete, so it contains partial data, but you get the gist.)
This can clearly be shared with the accountant, provided they have access to your workspace, which is easy to set up. It includes the number of documents, net, gross, and a "needs review" flag, with a few instructions on how to review, just by querying Receiptor through the MCP.
I've spoken with a lot of our customers one to one, and this is something they've been doing manually for a long time, especially those without an internal bookkeeper. If you think about how much time this used to take, it's extraordinary compared to the old way of doing things, and it just requires the connection to MCP.
I'll go back to the Calendly invoices to show that it extracted them with the description, the amount, and whether they were found as paid or not. We can modify these documents just by typing here, because we've made the tools available for Claude to call.
Recap of capabilities
To summarise: I've run a few queries to get answers back, and some queries to build custom dashboards and views of the data living in Receiptor. These can also be performed via the Receiptor agent on the platform, and via WhatsApp if the toggle is on for your number in the mobile scanners.
How to add Receiptor AI as an MCP connector in Claude
Now the simple process of adding Receiptor as an MCP connector to Claude. Go to Organization Settings, scroll down on the bottom left to Connectors, and add a connector. Choose custom connector, name it "Receiptor AI", and the link, which is also in the presentation we'll send to participants, is mcp.receiptor.ai.
I've already added it, so I won't add it again, but it will go through authentication and ask you to authorise it. Make sure you're authorising the correct workspace. After that you'll be able to see all the tools made available, though they're mostly technical. You can go to Customize Connectors (or Manage Connectors at the top left) to see the tools, for example from other MCPs, and to switch certain MCPs on and off from the chat. You can also explore everything just by typing "What can you do with the Receiptor MCP?"
Safety and limits
To wrap up, a quick word on safety and limits, because the question on people's minds is: do I really want an AI touching my financial data?
As I mentioned, Receiptor holds the data it has extracted from emails, WhatsApp and other sources. What this connection does is give read access to the AI so it can help you visualise that data. Any write operation requires your approval. You can automate the approval if you feel brave. Your credentials are never shared with the AI. The AI asks for confirmation, and you can connect and disconnect at any time, after which it no longer has access. It's still in the shape of an assistant, not an accountant, so it does require review and an eye on its output, but it's a real improvement on the previous manual work.
ChatGPT vs Claude
On setup, the connection in ChatGPT is similar to Claude, but it's less functional. OpenAI has done a lesser job of supporting MCPs: you can do it by creating a custom app, but the way Claude interfaces with MCPs is much better, so my recommendation is to use Claude for this type of thing.
Automations: weekly email to your accountant
One last thing worth showing. I instructed: "Use Receiptor automation to send a weekly email on Monday at 9am Pacific time to my accountant's email, with a CSV export of last week's paid documents." That's quite extraordinary, because it created an automation called "Weekly Paid Documents CSV Export" in Receiptor, which sends a CSV of the previous week's paid documents. So you can instruct Claude to run automations for you, sending CSV or PDF exports to email, and so on.
Native scheduling
Every query I've shown in this demo can also be scheduled natively via Claude. All the custom dashboards I've shown will appear as live artifacts, so they're all available here. The weekly automation can run as a native scheduled task in Claude, but in this case I specifically instructed it to run the automation inside Receiptor itself.
With that I'll stop, because it's a lot of information, but I'm happy to take questions.
Q&A: editing your own data
Question: Can Claude edit your own data in Receiptor?
Yes, it can. Let me show this on the full screen. Take the Calendly invoices this quarter: what's the category of these documents? They're already correctly categorised, but the other five documents are split between "software" and "web", which looks like inconsistent categorisation. So: "Categorise them all as General and Subscriptions." It will do that and ask me for confirmation, then actually modify the data in the system.
Claude wants to pull every document from Receiptor and put them all into General and Subscriptions, which is exactly what we want, so we'll accept the bulk edit and let it modify.
Q&A: proactive alerts on missing invoices
Question: How can the AI proactively alert us about missing invoices?
Yes, it can, through a scheduled task. I can set up a new task from here, or just ask Claude naturally to create a scheduled task from a standard chat. For example: "Every day at 9am Pacific time, alert me of any uncategorised expense or anything that might need attention."
In the meantime, the Calendly invoices for the quarter were updated and all five documents moved to General and Subscriptions. This could also have been done on the go via WhatsApp, but for the sake of this workshop we did it in front of you.
