Imagine asking an AI assistant "which customers have invoices more than 60 days overdue?" and getting the answer straight from your ERP, not from last week's spreadsheet. MCP (Model Context Protocol) handles that connection with a common standard, and Odoo 20, released in September 2026, already ships with its own MCP server.
In this guide you'll learn what MCP is, how it works, how it differs from an API, which tasks it can automate, what Odoo 20 exposes, and how to use it without granting more access than necessary.
What is MCP (Model Context Protocol) and why it became a standard
MCP (Model Context Protocol) is an open protocol that defines how an AI application connects to external systems, such as an ERP, a CRM, or a document repository, to query data and take actions. The system publishes its capabilities once and any compatible assistant can use them, with no separate integration for each one.
Anthropic introduced it in late 2024 and on December 9, 2025 donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. At that point there were more than 10,000 active public MCP servers and more than 97 million monthly downloads of its SDKs, and it was used by products such as ChatGPT, Gemini, Microsoft Copilot, and Visual Studio Code. That means connecting your ERP through MCP doesn't lock you into a single AI assistant.
How MCP works: host, client, server, tools, and resources
MCP has three participants: the host, the AI application the person uses (Claude, ChatGPT, or a code editor); the MCP client, which maintains the connection with each server; and the MCP server.
What is an MCP server
An MCP server is the program that exposes a system's capabilities, such as your ERP's, in a format any compatible client understands. It offers tools (actions, such as searching records or creating an activity), resources (context data, such as a database schema), and prompts (reusable templates).
The server sets the boundary: whatever it doesn't expose doesn't exist for the agent, so security is designed there, not in the prompt.
MCP vs API vs traditional integrations: when to use each
MCP doesn't replace APIs: an MCP server usually builds on them and adds a layer that AI can discover and use.
| API | Traditional integration | MCP server | |
|---|---|---|---|
| What it is | Interface for a program to read or write data | Programmed flow between systems | Standard layer that exposes tools to AI |
| Who decides | The code that calls it | Fixed rules | The model, within what's allowed |
| Best for | Custom development and high volume | Repetitive processes (order → invoice) | Queries that change case by case |
| Main risk | Code errors | Failures when exceptions occur | The model misinterpreting and acting on it |
If the process always follows the same steps, a workflow is cheaper and more predictable; MCP is useful when the question changes. In n8n, for example, a workflow can use MCP servers.
Data management tasks an agent can automate with MCP
To automate data management with AI, separate what the agent can do on its own from what a person must approve. From lowest to highest risk:
- Queries: "How much did we sell by product line this quarter?"
- Summaries of a customer, with open orders and outstanding invoices.
- Data quality: contacts with no email, possible duplicates.
- Suggested corrections that a person approves.
- Activities and notes in the CRM based on a conversation.
- Changes to prices, payment terms, or orders, always with human approval.
The first three are read-only, ideal for a pilot. If a rule is enough, don't use an agent: to get notified when a KPI goes out of range there are automatic threshold alerts, and to deal with duplicates, clean up your customer history.
MCP in Odoo 20: what it exposes by default and how to connect
Odoo 20 includes a native MCP server in its AI app. According to its official documentation, setup takes three steps:
- Create an API key scoped for MCP in your user preferences (Security tab) and choose how long it's valid.
- Configure the client with your database's endpoint (its URL followed by
/mcp) and the key in anAuthorization: Bearerheader. It requires Node.js, because it usesnpx -y mcp-remote; there are guides for Claude, Antigravity, and Codex. - Expose only what's needed: in developer mode, each additional server action is enabled with the "Available in MCP" option.
Odoo 20 MCP: the 5 tools enabled by default
By default, the Odoo 20 MCP server exposes five tools: Get Fields, Get Models, MCP Retrieve initial context, Search, and Read group. With them, the assistant learns the models and fields it can access, searches records, and retrieves grouped data, such as sales by salesperson. The rest, including those that write, such as Create Records and Update Records, stay hidden until you enable them.
Watch out for the "Readonly Tool" checkbox: it doesn't hide or block the tool. It only tells the client it can run it without asking for approval; checking it on an action that writes data is a risk.
Which Odoo plan you need
The MCP server documentation doesn't say which plan it requires. What is documented is that Odoo's external API is only available on the Custom plan, which is also where "Agentic AI" appears on the pricing page (checked in October 2026). Confirm this with Odoo or your partner; also, all of this applies only to databases already on Odoo 20.
Odoo 17, 18, and 19: third-party MCP and deprecated APIs
The Odoo 19.0 documentation doesn't include a native MCP server, so the alternative is third-party modules and servers, which we don't recommend without testing them. If you're evaluating a third-party MCP for Odoo, check:
- Which API it uses. Odoo has deprecated XML-RPC and JSON-RPC in favor of the JSON-2 API, and their
commonandobjectservices will be removed in Odoo 22 (fall 2028). - Who maintains it, what permissions it asks for, and which write tools it has enabled.
- Where your database runs. Odoo Online doesn't support custom modules with Python code, so an internal module requires Odoo.sh or on-premise, and an external server uses the API, which there is only available on the Custom plan.
If you're on Odoo 17, 18, or 19, the native MCP server is one more reason to upgrade your version.
MCP security: least privilege, human in the loop, and auditing
MCP security depends on which user the agent connects with, which tools you expose, and which approvals you require. In Odoo, every request is authenticated with the API key and the permissions of its user: the assistant sees what that user sees. Minimum checklist:
- Dedicated user with minimum permissions, never an administrator.
- Read-only at first; add write access one action at a time.
- Human confirmation for anything that writes.
- Short-lived keys, one per integration, as Odoo recommends.
- Control over what data leaves: whatever a tool returns is sent to the assistant's AI model, usually in its provider's cloud. If personal data is involved, read the guide to Chile's Law 21,719 on personal data protection.
- Logging: with a dedicated user, the agent's changes are recorded under its name.
How to start an MCP pilot
A good pilot answers a specific business question, such as letting the sales team look up their customers' accounts receivable in natural language, in read-only mode and with a dedicated user. Then you measure how many answers were correct and decide whether to add write access.
What about Autodesk Forma Data Management?
Forma Data Management is the common data environment of Autodesk Forma, the platform that now brings together Autodesk Construction Cloud; it used to be called Autodesk Docs. The official Autodesk MCP server catalog includes AutoCAD, Civil 3D, Fusion, and Revit, but not Forma Data Management (checked in October 2026). To bring progress billing and construction costs into an ERP, see Odoo for construction companies.
Why do it with KPB Solutions
Connecting an agent to your ERP is, above all, a matter of permissions and processes. As an Odoo Gold Partner with a team of more than 40 KPBeers, during your Odoo implementation we configure users, permissions, and which actions to expose.
With Kersify we build custom AI assistants and agents that query your CRM or ERP, whether it's Odoo or another system, with a least-privilege user, and hold every action that writes until a person approves it. If your database predates Odoo 20, we'll guide you through the version upgrade. And we practice what we preach: every guide on this site has a Markdown version for AI agents (add .md to its URL).
Want to connect AI to your ERP without granting more access than necessary? Book a meeting and let's design a focused, secure pilot together.
