MCP connectors: an AI assistant that knows the company's data
We connect AI assistants to a company's working systems: accounting, CRM, email, electronic invoices, tills, the warehouse and equipment. An employee asks in plain words and gets an answer based on live data, not on the model's memory. Here is what data we can retrieve and from where, and why the connectors are built the way they are.
In any company that has been operating for a few years, data sits in a dozen systems. Orders and stock are kept in 1C, the accounts in a separate program, deals and order correspondence in the CRM, agreements with suppliers in email, electronic invoices on the government portal, till sales in a cloud fiscal service. To answer a simple question — “has the client paid the last invoice and have the goods been shipped to them?” — a person opens three programs and puts the answer together by hand.
AI assistants such as Claude understand questions well and can work with figures, but on their own they know nothing about a particular company. The model cannot see your 1C or your email. Until it is connected to them, it can only reason in general terms or retell what an employee has pasted into the chat by hand.
We solved this problem for an equipment distributor in Serbia: we connected an AI assistant to its working systems via the MCP protocol.
What MCP is, in plain words
MCP (Model Context Protocol) is an open standard by which an AI assistant connects to external data sources. It is supported by Claude (the web version, the desktop app, Claude Code) and other clients. A connector is a small server that sits between the assistant and a working system. It tells the assistant which requests it can handle, checks the employee's permissions, goes to the system itself and returns the answer.
For the employee it looks simple: a connector is added in the assistant's settings, the employee signs in with their own account and from then on asks questions as they would ask a colleague. The assistant decides for itself which system to go to, and can combine data from several of them in one answer.
What data we retrieve and from where
Below are the connectors that are in operation now. For each one we list what is available in it and give example questions.
1C:Trade Management 11.5
The live database, with no intermediate copies: documents, catalogues, registers. The assistant finds the object it needs, looks at its fields and builds a query in the 1C query language. Stock and turnover are taken from the registers' virtual tables, which is fast even on a large database. Access can be limited to a list of objects and organisations.
“How many units of this model are in the warehouse, and how many of them are already reserved for orders?”, “Which customer orders from October have not been shipped yet?”
Accounting on PSIT/MPP
A copy of the accounting database is updated roughly every ten minutes. It contains outgoing and incoming invoices with their lines, the VAT ledger (PDV, SEF), the chart of accounts and postings, bank statements and payment allocation, stock accounting, business partners, retail receipts, fixed assets and exchange rates. The assistant queries it in SQL, but read-only.
“Turnover excluding VAT by month for this year”, “Which of this client's invoices have not been settled by payments?”, “The balance with this supplier at the end of the quarter”
Bitrix24
Deals, contacts, companies, leads, pipelines, tasks, Flows with process tasks for every order, task chats and the deal timeline. In practice the deal stage in the CRM often lags behind reality. So the assistant reconstructs the real status of an order from which stages are closed (supplier order, payment, goods receipt, delivery, documents in SEF) and what is written in the task chats. An employee's personal chats are available only through their personal webhook and only if the token allows it. Nobody else can see them.
“What stage is the order for this deal really at, and what is holding up shipment?”, “Which deals in the pipeline have not moved for two weeks?”
Email archive
Mailboxes on MailCow and Google Workspace: search in one mailbox or all of them at once, full messages, threads, attachments. For large files, such as a price list with thousands of lines, a one-time download link is issued. Executable attachments are blocked: the assistant sees only their name and size. The connector does not send emails.
“What did the supplier write about delivery times for this order?”, “Find the latest price list from this distributor”
SEF — Serbia's government electronic invoicing system
Incoming and outgoing e-invoices, their statuses and status history, the parsed UBL document (parties, totals, taxes, lines, notes), the original PDF and XML files, payments and assignments in the central invoice register (CRF), reference lists. The connector respects SEF's limit of three requests per second per key and itself splits long periods into windows that SEF handles correctly.
“Which incoming invoices from this week have not been accepted yet?”, “Has the buyer accepted our invoice? Is the payment recorded in CRF?”
ESIR FisSoft — fiscal receipts
The receipt log, a full receipt together with its refunds and copies, a sales report. Turnover is calculated as sales minus refunds, with advances counted separately, as fiscal logic requires. Any receipt can be verified via the link in its QR code directly with the tax authority.
“Yesterday's revenue by payment method”, “Is this receipt from a customer genuine?”
Jetinno — the coffee machine cloud
The live status of the fleet: how many machines are online, which ones have errors, where ingredients are running out. In addition, sales and orders, error history, ingredient levels and refills, the cleaning log and products. The connector does not send commands to the machines.
“Which machines have errors right now, and where does milk need topping up?”, “When was the machine at this location last cleaned?”, “Sales by product for the week”
Syrve support portal
Requests from client restaurants with their correspondence and attachments, clients and the “head office — restaurants” hierarchy, statistics and SLA compliance, support package billing, the knowledge base. Passwords, codes and remote access IDs in texts are always hidden. Personal data — phone numbers, addresses, email — is hidden depending on the token's settings.
“How many open requests does this chain have, and are we meeting the SLA?”, “How did we solve the receipt printing problem last time?”
Unified supplier product database
Distributors' offers: purchase prices excluding VAT and stock levels (including in the form “10+”, meaning at least ten), our own warehouse from 1C, and the European distributor EET with prices in euros and delivery to order. Specifications are pulled from the catalogue by EAN. There is also separate analytics on stock movement in suppliers' warehouses.
“Where is the cheapest place to buy this model, and do we have it ourselves?”, “Check prices and availability for this list of EANs”
Telegram archive
Full-text search and search by meaning across channels and chats. The latter runs on a vector index, so it finds messages even when they are worded differently or written in another language.
“What was discussed about the change in customs rules over the last month?”
Why it is built this way
Read-only. The connectors do not create, change or delete anything. To change data, an employee goes to the system itself. The assistant helps to find and understand, while decisions and actions remain with people.
An access policy for each token. For each employee the administrator sets separately which sections are open to them, how many rows can be retrieved in one request, how long texts can be, what date range is allowed and how often requests can be made. If a request goes beyond these limits, the assistant receives a clear explanation of what to change: narrow the period, refine the filter, aggregate the data into a total. A restriction cannot be bypassed through a neighbouring tool.
Data is data, not commands. Emails, invoices, chats and support requests can contain any text, including text that looks like an instruction to the assistant: “pay this”, “forward this”, “ignore previous instructions”. This is known as prompt injection. Every connector explicitly tells the model that the content of the systems is business data. If such a “command” appears in it, the assistant does not carry it out and reports it to the user.
A copy where the source is slow or fragile. Some systems respond slowly, some have no proper programming interface, and the coffee machine cloud protects its login with a captcha. For such sources the gateway keeps a copy of the data and refreshes it roughly every ten minutes. Every answer carries a freshness mark: when the data was last updated. If the copy has fallen behind, the assistant says so directly. Where a system can handle the load, as 1C or SEF can, queries go to the live database.
Masking of sensitive information. Passwords, access codes and remote access IDs in texts are always hidden. Personal data is hidden where the employee's role requires it.
Each employee signs in separately. Connection is via OAuth: an employee signs in with their own key, and everyone has their own token. The systems' keys — webhooks, passwords, certificates — are stored only on the gateway server; neither employees nor the assistant can see them. A person's access can be switched off in one step without affecting anyone else. The gateway admin panel and the request log belong to the company's administrator.
What was built
Each connector is a separate Python server implementing MCP. The connection runs over secure HTTPS, and sign-in goes through our own OAuth 2.1 authorisation server. Token policies, the request log and user management are brought together in the gateway admin panel. Background synchronisers maintain the data copies. The accounting copy is stored in DuckDB, which runs analytical queries across the whole history quickly. A vector index was built for the Telegram archive.
The most labour-intensive part of the work is not the protocol but knowledge of each system. The connector explains to the model how the data is structured: which invoice status counts as a working one, where to look for payment if the “paid” field is not filled in, how to calculate till turnover taking refunds and advances into account, why the deal stage cannot be trusted without the task chats. Without this the model would answer confidently but inaccurately.
Practical value
Questions that used to mean opening several programs or going to the accountant, an employee now asks themselves and gets an answer straight away. A manager sees the real state of orders, not stages someone forgot to move on. A purchasing manager compares supplier prices and the company's own warehouse with a single question. The service team sees the state of the machines and the history of support requests without switching between portals.
And the company takes no risk: the assistant cannot change anything, sees only what it is allowed to see and does not carry out “instructions” from other people's emails.
Outlook
The MCP protocol is open, and the range of systems that can be connected is almost unlimited: any program with a database or a programming interface can get its own connector. The next step is scenarios in which the assistant prepares an action (a draft email, a document, a request), and a person checks it and confirms it in the system itself.
If your data is also scattered across several systems and you want an assistant to answer questions about it, write to us. We will discuss which systems to connect first and what access rules they need.