“Can ChatGPT look directly into our Odoo?” Yes — via a remote MCP server, and with the same restriction that applies to all clients: the AI only sees what you release. Below is what works today and how to set it up.
What you need
- A hosted MCP server with a publicly reachable HTTPS endpoint. Clients running in the cloud cannot reach an MCP server on your laptop.
- Authentication that the client supports: a bearer token, or OAuth 2.1 with discovery and dynamic client registration. At DAADit both are available; the OAuth metadata lives at the standard
/.well-known/paths, so clients find the authorisation server themselves. - An Odoo user with an API key (not a password) and exactly the permissions the use case needs.
Per client, in brief
- Claude (web and desktop): connectors with remote MCP servers and OAuth. In practice the most mature experience for long analysis sessions.
- ChatGPT: connecting external MCP servers is available in the business/developer context; the exact place in the interface and availability per subscription change regularly — check this in your own account before you promise it internally.
- Mistral Le Chat: supports remote MCP connectors; attractive for European organisations that want to keep their model choice within the EU.
- Cursor, Continue, Goose: MCP is built into the developer tooling and is ideal for quickly testing what your toolset actually returns.
The configuration in practice
At DAADit you create a connection in the customer portal with four details: your Odoo URL, the database name, the username and an API key. You choose the Odoo version (16 to 19) and which actions are allowed; read-only by default. You then receive an endpoint of the form https://daadit.odoo.com/mcp/i/<your-slug> plus a key that you get to see only once. You put those two into your AI client. From that moment on, the client receives only the tools you have released on tools/list.
What to test in the first session
- Ask for the models the client can actually see — does that list match what you intended?
- Ask a question that combines several models (for example, open quotes from customers with a recent support ticket). That shows whether your toolset is sufficient.
- Explicitly try something that is not allowed (changing a record). You should get a polite refusal, not a silent failure.
What you will run into
- Context limit. Large result sets fill up the window. Limit fields and record counts; let the model filter via the tool, not in its head.
- Tool descriptions. The model chooses based on text. Vague descriptions lead to the wrong tools and pointless calls.
- Differences between clients. Each client has its own limits, timeouts and behaviour with parallel calls. Test the client your team actually uses.
- Changing support. This market moves month by month. So don't lock your process into a single client.
The order we recommend
Start read-only on staging, prove one concrete question, then apply the same scope to production and only expand once the logging shows which tools are actually used. That keeps the discussion about value, not about risk.
Practical tip: record one test script per client (the same three questions from the section above) and run it again after every client update — that way you spot behavioural changes before your users call about them. Unsure which client suits your team, or stuck on the OAuth configuration? Contact DAADit Group — we are happy to take a look with you.
Verder lezen: Connecting Claude to Odoo and MCP security best practices.
Want to try it yourself? The DAADit Hosted MCP server for Odoo is free to start with 50 calls per day — no installation, no credit card. Create a free account or read the technische productdocumentatie.