How to give an AI agent access to 1C, Odoo or Bitrix24 without breaking your data

To connect an AI agent to 1C, Odoo or Bitrix24, you do not give it the database password or access to the whole API. You give it an MCP server: a small service that exposes a few predefined actions, such as finding a counterparty, showing an order or checking stock.

Each action validates its input and limits the response size and execution time, and a single setting forbids any data changes. A simple read-only server with three actions, my Basic package, takes 5 days and costs $400.

What MCP is, in plain words

MCP, the Model Context Protocol, is an open standard for connecting AI agents to external systems. Since December 2025 it has been governed by the Agentic AI Foundation under the Linux Foundation, so no single company owns it.

The server describes which actions are available and what parameters they take. The agent picks the right action for the employee's question, and the server runs it and returns the result.

One server works with any program that supports MCP: agents in code editors such as Cursor and VS Code, a company's own agents, and ChatGPT. There is no need to rewrite it for each program.

ChatGPT has its own conditions: the server is added in developer mode on the web, on Plus, Pro, Business, Enterprise and Edu plans. ChatGPT reaches the server over the network, so the server has to be available at a public address and protected with authorization.

Why an agent should never get the database password

An agent makes mistakes the way a new employee does: it misreads a question and changes the wrong record. The difference is that with full API access it can change or delete a thousand records in one request.

There are two simpler problems too. Without a response limit, the agent pulls thousands of rows and hits the model's limits. And a slow 1C response freezes the conversation, while the employee has no idea why nothing is happening.

How safe access is built

• Actions follow the business process, not database tables. Creating an order for an existing counterparty is its own action with its own checks. There is no universal "run any query" action.
• Input validation. A wrong parameter is rejected before the system is called.
• Limited responses. If there is more data than the limit, the agent gets part of it and a note that the response was cut.
• Read-only mode. One setting forbids any changes.
• Time limits. If the system responds slowly, the request is stopped and the agent tells the employee.
• Write confirmation. Actions that change data run only after a person confirms them.

MCP does not make the agent smarter. Given a vague question, it can still pick the wrong action. The server is there for something else: so that such a mistake costs a wrong answer in the chat, not damaged data in 1C.

Example: actions for a sales team

Action	What it does	Changes data
Find counterparty	search by name, tax ID or phone	no
Customer orders	recent orders and their statuses	no
Check stock	product stock by warehouse	no
Unpaid invoices	overdue invoices for a customer	no
Create order	new order for a found counterparty	yes, with confirmation

A manager asks: "How many units of SKU 8812 are in stock, and does Romashka LLP have unpaid invoices?" The agent calls two actions and answers in one message. Nobody has to search through different 1C windows.

A set of five actions like this, one of which changes data, matches the Standard package. Basic includes up to three read-only actions.

An open-source MCP server for Odoo

You can see how this looks in code in my open-source MCP server for Odoo. It has six actions: counterparty search, order list and details, order creation and confirmation, stock, and invoice list.

Input is validated, changes are switched off with one setting, and every request has a time limit. The server connects over stdio on a single computer or over HTTP on a shared server.

An MCP server for 1C and Bitrix24

For 1C and Bitrix24 the server is built the same way, only the connection differs. It talks to 1C through the standard OData interface built into the platform, or through HTTP services when some checks have to run inside 1C. It connects to Bitrix24 through its REST API.

If 1C runs on your own server, the MCP server can sit next to it. The database is not copied anywhere: the model's chat receives only the answer to a specific question.

Timeline and price

Package	Includes	Timeline	Price
Basic	up to 3 read-only actions in one system, Docker image, setup instructions	5 days	$400
Standard	up to 8 actions, including data changes with confirmation, automated tests	10 days	$900
Premium	everything in Standard, authorization for HTTP connections, installation on your server and handover	15 days	$1,600

More on the MCP servers for AI agents page.

Where to start

• Write down 5–10 questions your employees would ask the agent.
• Mark which of them only read data and which change something.
• Check where the system runs, in the cloud or on your server, and whether it has an API.
• Start in read-only mode: it is useful right away and cannot break anything.

With this list you can send me the task. I will reply with a plan, the main risk and a fixed price.