Summary: Your Odoo is now an MCP server on its own — AI assistants like Claude, ChatGPT and Gemini connect straight to your Odoo URL and sign in with their Odoo account, with per-user permissions, read-only access and a full audit trail.

Previously, connecting an AI assistant meant installing and running a separate bridge program on your machine. Now Odoo speaks the assistant's language itself. You point the assistant at your Odoo address and it connects. Nothing to install, nothing to keep running.

Instead of copying an API key into your AI assistant, you can now simply log in. The assistant sends you to a normal Odoo login page, you approve the connection once, and you are done. Each person connects as themselves, so everyone sees exactly the data their Odoo user is allowed to see — no shared key, no shared identity.

For setups where logging in through a browser is not an option — a server, an automated script, or an AI tool that does not support the login flow — you can still authenticate by pasting an Odoo API key into the tool's authorisation header. If a tool connects without one, Odoo now tells it where to go and log in, so most modern assistants pick up the login flow by themselves.

On the approval screen there is a checkbox: "Allow creating and modifying data". Leave it unticked and the assistant can look but not touch — it cannot create, edit or delete anything, and it is not even shown the tools that would let it try. Ideal for letting a colleague ask questions of your data without any risk to it.

When you create an API key you can now mark it as MCP-only. Such a key works for AI assistants and nothing else, so if it ever leaks it cannot be used to get at the rest of your Odoo.

Rather than handing an assistant broad permission to create and edit records, an administrator can now publish specific, named actions — "confirm this sales order", for example. The assistant sees a short list of approved things it can do, and each one still runs under the permissions of the person asking.

On connecting, the assistant is told your name, your timezone and which company you are working in. That means dates land on the right day and records land in the right company, instead of the assistant guessing.

When you ask about a record, the assistant now sees its linked items — order lines, contacts, tasks — either summarised as a count or previewed by name, rather than as a meaningless list of internal numbers.

But Also...

- A complete activity log: every request now records who asked, from where, how they signed in, which assistant they used, what was accessed and how much came back. All of it is visible under the MCP Server menu.

- New shortcuts in Settings take you straight to connected AI apps, active sessions, and your published actions.

- The documentation has been rewritten around the three ways to connect, with a table of which AI assistants have been tested and confirmed working.

- Guidance for running this across multiple databases on one server.

- Fixed a crash on the approval screen: on any Odoo instance that also runs a website, clicking through to approve an AI assistant returned an error page instead of the approval form. Approving a connection now works everywhere.

- Closed a gap where an assistant could read the structure of a model — its fields and layout — that it had not been granted permission to read.

- Fixed a case where a request with an unusually long name could disappear from the activity log entirely. Every request is now recorded.

- Error messages no longer leak internal database details, such as the contents of a conflicting record, back to the assistant.

- Assistants can no longer trigger scheduled jobs or server actions as a back door to elevated permissions.

- Passwords, tokens and secret keys are now filtered out of bulk record listings as an extra safety net.

- A request without a limit no longer pulls an entire table; results are capped at your configured maximum.

- Turning the MCP server off, or changing the allowed browser origins, now takes effect immediately instead of up to five minutes later.

- Oversized requests are rejected up front rather than being read into memory, and repeated failed sign-in attempts can no longer be used to hammer the server.

- Numerous stability fixes so that a busy database, a malformed request or a failed log write can no longer turn a successful action into an error — or quietly undo it.

Summary: The full AI-assistant experience from Odoo 19 now runs on Odoo 18 — assistants like Claude, ChatGPT and Gemini connect straight to your Odoo URL and sign in with their Odoo account, with per-user permissions, read-only access and a full audit trail.

Previously, connecting an AI assistant meant installing and running a separate bridge program on your machine. Now Odoo speaks the assistant's language itself. You point the assistant at your Odoo address and it connects. Nothing to install, nothing to keep running.

Instead of copying an API key into your AI assistant, you can now simply log in. The assistant sends you to a normal Odoo login page, you approve the connection once, and you are done. Each person connects as themselves, so everyone sees exactly the data their Odoo user is allowed to see — no shared key, no shared identity.

For setups where logging in through a browser is not an option — a server, an automated script, or an AI tool that does not support the login flow — you can still authenticate by pasting an Odoo API key into the tool's authorisation header. If a tool connects without one, Odoo now tells it where to go and log in, so most modern assistants pick up the login flow by themselves.

On the approval screen there is a checkbox: "Allow creating and modifying data". Leave it unticked and the assistant can look but not touch — it cannot create, edit or delete anything, and it is not even shown the tools that would let it try. Ideal for letting a colleague ask questions of your data without any risk to it.

When you create an API key you can now mark it as MCP-only. Such a key works for AI assistants and nothing else, so if it ever leaks it cannot be used to get at the rest of your Odoo.

Rather than handing an assistant broad permission to create and edit records, an administrator can now publish specific, named actions — "confirm this sales order", for example. The assistant sees a short list of approved things it can do, and each one still runs under the permissions of the person asking.

On connecting, the assistant is told your name, your timezone and which company you are working in. That means dates land on the right day and records land in the right company, instead of the assistant guessing.

When you ask about a record, the assistant now sees its linked items — order lines, contacts, tasks — either summarised as a count or previewed by name, rather than as a meaningless list of internal numbers.

But Also...

- A complete activity log: every request now records who asked, from where, how they signed in, which assistant they used, what was accessed and how much came back. All of it is visible under the MCP Server menu.

- New shortcuts in Settings take you straight to connected AI apps, active sessions, and your published actions.

- The documentation has been rewritten around the three ways to connect, with a table of which AI assistants have been tested and confirmed working.

- Guidance for running this across multiple databases on one server. Note that on Odoo 18, multi-database instances need to select the database by web address — the header-based method is an Odoo 19 feature.

- Everything was re-verified on a fresh Odoo 18 database, with the full test suite passing and a live connection confirmed end to end.

- Closed a gap where an assistant could read the structure of a model — its fields and layout — that it had not been granted permission to read.

- Fixed a case where a request with an unusually long name could disappear from the activity log entirely. Every request is now recorded.

- The approval screen displays correctly on Odoo instances that also run a website, instead of returning an error page.

- Error messages no longer leak internal database details, such as the contents of a conflicting record, back to the assistant.

- Assistants can no longer trigger scheduled jobs or server actions as a back door to elevated permissions.

- Passwords, tokens and secret keys are now filtered out of bulk record listings as an extra safety net.

- A request without a limit no longer pulls an entire table; results are capped at your configured maximum.

- Turning the MCP server off, or changing the allowed browser origins, now takes effect immediately instead of up to five minutes later.

- Oversized requests are rejected up front rather than being read into memory, and repeated failed sign-in attempts can no longer be used to hammer the server.

- Numerous stability fixes so that a busy database, a malformed request or a failed log write can no longer turn a successful action into an error — or quietly undo it.