Use MCP to connect AI to Odoo when the work is open-ended: questions that span sales, accounting and inventory, or small actions a person decides on in conversation. Use a custom API integration when the job is fixed and repetitive. Use Odoo's built-in AI app when people work inside Odoo screens on Odoo 19 or 20 Enterprise. Most companies end up with a mix, and the right choice depends on who asks and where.
Three ways to put AI on Odoo data
| Custom API integration | Odoo built-in AI (AI app) | Odoo MCP server | |
|---|---|---|---|
| Where the user works | Nowhere, it runs in the background | Inside Odoo | In their AI assistant |
| Who decides what runs | The developer, in code | Odoo agent configuration | The model picks tools; the server limits them |
| Odoo versions | Any with external API | 19+ Enterprise | 16+ via external API; native in Odoo 20 |
| AI model | Optional | OpenAI or Gemini via Odoo | Whatever the AI client uses |
| Good for | Fixed, high-volume jobs | Drafting text, AI fields, document sorting inside Odoo | Cross-app questions, managers, small confirmed actions |
| Weak for | Open questions | Users outside Odoo, older versions | Bulk changes, strict real-time sync |
When a custom integration is the better choice
A webshop order sync, a nightly export to a BI tool or an EDI exchange should be plain code. It is predictable, testable and cheap per call, and it does not need a language model to decide anything. Adding MCP there adds cost and uncertainty with no benefit.
Where custom code struggles is the long tail: "which customers bought less this quarter than last", "what do we owe supplier X and is anything overdue", "draft a quote like the one we sent last month". Each of these would need its own report or endpoint. An MCP server with a handful of read tools answers all of them.
When Odoo's built-in AI is enough
Odoo 19 introduced an AI app with agents, AI fields, AI server actions and text drafting in the editor. If your team lives in Odoo screens and runs Odoo 19 Enterprise on the Custom plan, start there: drafting e-mails, filling a classification field or sorting documents does not need an external assistant.
Its limits matter for some companies:
- it is not available in Odoo Community or in Odoo 16 to 18;
- the standard Ask AI agent cannot create or change records, according to Odoo's documentation, so actions need configured agents and server actions;
- it serves Odoo users, not a manager or partner who works mostly outside Odoo.
Odoo 20 adds its own MCP server, which shows that Odoo sees external assistants as a complement to the in-app AI, not a replacement.
What MCP adds
One connection, several assistants
The same server works with any client that supports the MCP specification. Switching assistants does not mean rebuilding the integration.
Synthesis across apps
The model calls several tools in sequence and combines the results. That is where it saves the most time compared with opening several Odoo menus.
Access without Odoo screens
A manager asks in their assistant and gets an answer from live data, filtered by their own Odoo access rights.
Works on older versions
Odoo 16 to 19 and Community can use MCP through the external API, without an upgrade project.
What MCP does not solve
- Data quality. Duplicate partners and missing product categories produce poor answers through any interface.
- Correctness. The model can misread a question or a field. Ask it to cite record references and spot-check against Odoo reports.
- Bulk work. Updating hundreds of records through chat is slow to review. Use imports or server actions.
- Security by default. A server that connects as admin and exposes generic writes is a liability. The controls on the security page are what make MCP usable on production data.
A simple decision rule
- If no person is in the loop, write integration code.
- If people work in Odoo screens on 19 or 20 Enterprise, use the AI app first.
- If people ask cross-app questions from outside Odoo, or you run 16 to 19 or Community, use an MCP server with least-privilege users and confirmed writes.
Concrete prompts per role are listed on Odoo MCP use cases.
Connect AI to your Odoo
Describe your situation in a few lines. We reply by e-mail.
Frequently asked questions
Is a custom integration ever better than MCP?
Yes. A fixed, high-volume job such as syncing orders from a webshop every five minutes should be ordinary integration code with no AI model in the loop. MCP pays off for open-ended questions and occasional actions that a person starts from a chat.
Do we still need MCP if we use Odoo 19's AI app?
Not necessarily. The AI app works inside Odoo screens and is an Enterprise feature. MCP adds value if people want answers in their own AI assistant, if managers rarely open Odoo, or if you run Odoo 16 to 18 or Community, where the AI app does not exist.
Does MCP lock us into one AI vendor?
No. MCP is an open specification supported by several AI clients, so the same Odoo MCP server can serve Claude, ChatGPT, Cursor and others. The clients still differ in plan limits and in how they show confirmations for write actions.
What does MCP cost to run?
The protocol is free. Costs come from the AI client subscription or API usage, from hosting or subscribing to an MCP server, and from Odoo itself, since external API access on Odoo Online requires the Custom plan. Each of these is priced by its own vendor.