← Back to Blog
Engineering by Featured

Odoo's Own MCP Server: Six Real Requests and What It Still Lacks

Odoo's Own MCP Server: Six Real Requests and What It Still Lacks

Summary

Odoo 20 Enterprise ships an MCPAn open protocol developed by Anthropic that standardizes how AI models connect to external tools, data sources, and services. MCP allows LLMs to call tools (file systems, APIs, databases) in a… server with OAuth sign-in. I connected Claude Code as an admin and as a sales user and ran six real requests. It reads safely within access rights; write limits and the audit trail are thin.

In April, Connecting AI Agents to Odoo with MCP compared the four MCP servers for Odoo that existed at the time and picked muk_mcp as the place to start. It ended on a gap none of them had closed: there was no OAuth flow. Every user had to generate an API key by hand and paste it into their AI client.

Seven months later, Odoo S.A. ships its own MCP server as part of the AI app in Odoo 20 Enterprise, and it signs users in with OAuth. The community options moved too; Connecting AI Agents to Odoo with MCP, Seven Months Later covers them and what we would pick now.

I set the official server up on a local Odoo 20 database, connected Claude Code to it as two different users, and ran six real requests. This post shows the setup screens, what the agent did and answered, and where the server fell short.

Short answer: if you run Odoo 20 Enterprise (online or self-hosted), you can turn it on for reading. The AI sees only what the person who signed in can already see. Keep the tools that create or change records off until you have a clear reason and a small group of users for them. Business readers can read the first sections and the six requests, then jump to Should you turn it on?

Before you start

MCP lets an AI app call Odoo’s functions, such as “search records”, as you; the first post covers who uses it and why. Here is what happens on each question:

flowchart LR
  U["You ask a question<br/>in Claude, ChatGPT, Codex..."] --> C["AI client"]
  C -->|"MCP tool call,<br/>signed with your token"| O["Odoo<br/>/mcp endpoint"]
  O --> R{"Your access rights<br/>and record rules"}
  R -->|allowed| D[("Odoo data")]
  R -->|refused| X["Error message"]
  D -->|result| C
  X -->|result| C
  C -->|answer| U

The AI never gets a database password. It can only use the tools Odoo exposes, with the rights of the person who signed in. Two more things to know:

  • What you need: Odoo 20 Enterprise, online or self-hosted (not Community, and not 19.0 or older), and an AI app that lets you add a custom MCP server or connector, such as Claude, ChatGPT or Claude Code.
  • Where the data goes: every record the AI reads is sent to the AI provider (Anthropic, OpenAI…) as part of the conversation, the same way a pasted spreadsheet would be. Check that this fits your company’s policy before you connect real data.

What MCP is, and what it is not

  • A shared language, not a tunnel. Odoo publishes a menu of tools. The AI app can call those tools and nothing else; it never touches the database directly.
  • The AI app calls Odoo, not the other way round. The language model runs at the AI provider, and every record it reads is sent there.
  • It works as you. You sign in to Odoo once and click Allow. Every call then has your access rights, no more.
  • It replaces menus for one-off questions. Instead of opening the invoice list and filtering, you ask "which invoices are overdue?" and get a table in under a minute (request 1). It also lets automation tools such as n8n work with Odoo. Routine reports are still quicker from the menu.
  • It is not web search, and it is not always right. It answers from your Odoo data, can only use the tools your admin turned on, and can miss what a person would catch (request 4).

A few terms come up again below:

Tool One function the AI can call, for example search, group totals or create records.
OAuth "Sign in with Odoo": you log in once in the browser and approve the AI app, instead of copying a key.
API key, token A long secret string that stands in for your password. OAuth creates one for you; Odoo shows it in your preferences, where you can delete it.
Access rights Which kinds of records a user may read, create, edit or delete (invoices, quotations, contacts...).
Record rules Which records of that kind they see, for example "only my own quotations".

Odoo’s official MCP server

What it is

The module is ai_mcp, in Odoo Enterprise. A first version came with Odoo Online 19.4, with API keys only. OAuth sign-in, which ChatGPT and Claude need, arrived in 20.0, the version this post covers. It is documented in Odoo’s MCP server guide1. It depends on the ai module and installs itself when the AI app is installed. The documentation pages do not exist for 19.0 or earlier.

Your database gets one address for AI clients, https://<your-database>/mcp. Behind it are the AI tools that Odoo’s own AI agents already use: search, group totals, list models, create and update records, open a view, read accounting reports, edit a website page, and more. An external AI app calls the same code that Odoo’s built-in agents do.

Two engineers from Odoo’s AI team explained how they built it in a talk published on 30 September 2026, Odoo MCP: A technical deep-dive into AI connectors6. I refer to it below as “the talk”. It runs 55 minutes; the parts this post draws on most are why the server is stateless (14:04), building and testing a tool (27:13), connecting n8n, Claude and ChatGPT (36:11) and the Q&A (49:03).

Odoo MCP: A technical deep-dive into AI connectors, by Odoo's AI team (30 September 2026, 55 minutes).

Their rule of thumb: MCP is for automations and outside apps, such as n8n or ChatGPT, and Odoo’s built-in AI is for work inside Odoo. The built-in AI chat uses the same tools but does not go through MCP.

The talk also answers why Odoo added MCP when it already had an API (JSON-RPC). With the API, a developer or an agent has to learn the endpoints from the documentation first, and an agent that reads whole API docs fills its context and costs more per question. With MCP, the agent only loads a short description of each tool and reads the details when it decides to use one. Any AI app that speaks MCP can connect without custom code.

Their n8n demo is a good example of an automation. An n8n workflow picks up a customer email that asks where an order is. An AI agent in the workflow looks the order up in Odoo through MCP and replies. When it cannot answer from the data, as with a question about a broken item, it opens a helpdesk ticket with the customer’s email attached and a reminder for the person assigned. In the demo the agent could not confirm the order’s status, so it took the second path and told the customer a person would follow up.

Setting it up, step by step

The test database ran Odoo 20.0 Enterprise (commit 111432a1, 6 November 2026) with demo data, on my laptop. Steps 1 to 3 are for your Odoo admin. Steps 4 to 6 are for each person who wants to use an AI app with Odoo.

Step 1. Admin: install the AI app

ai_mcp comes with it; there is nothing else to install.

Step 2. Admin: check which AI apps may sign in

Under Settings → General Settings → Integrations → MCP OAuth, Odoo keeps a list of AI apps allowed to sign in with OAuth. Out of the box it lists ChatGPT, Claude (web and desktop), Claude Code and Grok. Any other client, Codex for example, must be added here by its metadata URL, or use an API key. The Allow Dynamic Client Registration box (DCR) lets any app register itself. It is off by default because, as Odoo’s team put it in the talk, anyone could otherwise flood the registration endpoint. For a client that is not on the list, such as n8n in their demo, either add its metadata URL or tick DCR, connect the client, and untick it again.

Odoo settings, MCP OAuth block: four allowed client metadata URLs for ChatGPT, Claude, Claude Code and Grok, and an unticked option to allow dynamic client registration
Settings → Integrations → MCP OAuth: the AI apps allowed to sign in, and the DCR option, off by default.

ChatGPT and Claude (web and desktop) connect from their own servers, so your Odoo must be reachable from the internet over HTTPS. Claude Code and other apps running on your computer can also reach an internal Odoo, but Odoo itself still needs outbound HTTPS to fetch each app’s sign-in details (see the developer notes).

Step 3. Admin: choose the tools

Each AI tool is a server action with an Available in MCP checkbox (Settings → Technical → Server Actions, in developer mode). Five reading tools are ticked by default. Odoo’s team keeps the write tools off on purpose: a destructive call made by mistake is expensive to undo, and some tools only navigate Odoo’s screens, which an outside agent has no use for. In the screenshot, the reading tools Search and Read group are further down the list. I had also ticked Create Records for the write tests in this post.

List of AI Tool server actions in Odoo with the Available in MCP and Readonly Tool columns; Create Records (ticked for the tests), Get Fields, Get Models and MCP Retrieve initial context are ticked
The AI tools are server actions. Available in MCP exposes a tool; Readonly Tool is only a hint for the AI app.

Readonly Tool is a label for the AI app: tools marked this way may run without asking you, and any other tool should make the app ask first. Odoo itself does not enforce it.

To keep a tool for some users only, open its server action and fill Allowed Groups (at the top of the form, in developer mode). I limited Create Records to Sales / Administrator and checked it with both users: Mitchell still had six tools and created a contact; Sam’s AI app listed only the five reading tools, and calling Create Records directly anyway got “You don’t have enough access rights to run this action.” I ran this check after the six requests below, so in request 4 Sam still had Create Records.

This group check is about the tool. It comes before Odoo’s usual access rights on the data, which still apply to everyone who can use the tool, as in request 5.

Odoo server action AI Tool: Create Records with Allowed Groups set to Sales / Administrator
Allowed Groups on Create Records: only Sales administrators can see and use the tool.

Step 4. User: add Odoo to your AI app

For Claude Code it is one command:

claude mcp add --transport http odoo https://<your-database>/mcp

The talk showed the other two apps, which I did not test:

  • Claude: Customize → Connectors → Add custom connector, then paste https://<your-database>/mcp.
  • ChatGPT: first turn on developer mode (Settings → Security and login), then Plugins → + → Create app → Create MCP app with the same address.

Step 5. User: sign in and approve

The first time the app calls Odoo, your browser opens Odoo’s login page. You log in as usual and see this:

Odoo Grant Access page: Claude Code is requesting access to your Odoo account for MCP; access expires after 1 Day, 1 Month (selected), 1 Year or Never; Allow and Deny buttons
The consent page Odoo shows when Claude Code signs in. You choose how long the access lasts.

You choose how long the access lasts. The choices depend on your groups: admins with Settings rights also get Never, while a regular user’s keys are capped at 90 days, so they see 1 Day and 1 Month. Admins should avoid Never: with no refresh tokens and no audit trail, a key that never expires is the riskiest choice. After Allow, the app can work with Odoo until the access expires, then it asks you to log in again.

Step 6. User: see or remove the access

The access shows up as an API key named after the app, under My Preferences → Security. Delete removes it at once; the app’s next request is refused.

My Preferences, Security tab: API keys list with mcp test admin (MCP), rpc test admin (RPC) and Claude Code (MCP), each with created and expiry dates and a Delete button
Each connection is an API key under My Preferences → Security. Delete revokes it at once.

For developers: the sign-in flow and self-hosting

This is the full exchange behind steps 4 to 6, for Claude Code. ChatGPT and Claude (web and desktop) receive the code on their own servers instead of a local callback; the rest is the same.

sequenceDiagram
  participant You
  participant Client as Claude Code
  participant Browser
  participant Odoo
  Client->>Odoo: call /mcp (no token yet)
  Odoo-->>Client: 401 + where to sign in
  Client->>Browser: open Odoo's sign-in page
  You->>Browser: log in, choose 1 month, click Allow
  Browser->>Odoo: consent
  Odoo-->>Browser: redirect with a one-time code
  Browser-->>Client: code, on the client's local callback
  Client->>Odoo: code + PKCE proof
  Odoo-->>Client: access token (scope "mcp")
  Client->>Odoo: tool calls with the token

I drove this flow with a script, using Claude Code’s published client ID (https://claude.ai/oauth/claude-code-client-metadata) and a loopback callback, the way Claude Code does it. Odoo fetched the client’s metadata, showed the consent page above, and issued a token with scope mcp, valid for the month I picked. With that token, tools/list returned the same tools as with an API key. Details that matter when you integrate or self-host:

  • It is the authorization code flow with PKCE (S256) for public clients, so no client secret sits on the user’s laptop. Clients identify themselves with a client ID metadata document, as the 2025-11-25 MCP spec expects.

  • There are no refresh tokens. The token is a regular Odoo API key with scope mcp, named after the client.

  • OAuth is refused for portal users; they need an API key with the MCP scope. An RPC key is refused on /mcp. To use a key with Claude Code:

    claude mcp add --transport http odoo https://<your-database>/mcp \
      --header "Authorization: Bearer $ODOO_MCP_KEY"
    
  • Odoo builds every sign-in URL from web.base.url. Behind a reverse proxy, run with proxy_mode and make sure that value is the public HTTPS address, or clients get the wrong URLs. On a multi-database server, the host must map to one database (dbfilter).

  • Odoo downloads the client’s metadata document on each authorize and token request (5 second timeout, no redirects), so the server needs outbound HTTPS to claude.ai, chatgpt.com and the others.

  • Removing a client from the allowlist only blocks new sign-ins. Keys it already received keep working until they expire or are deleted.

Three design choices that the talk explains:

  • Stateless, on purpose. A stateful MCP server keeps a streaming (SSE) connection open per client. One Odoo server often serves thousands of databases with a few synchronous workers, so a handful of connected AI apps could block every worker. Odoo tools finish quickly, so there is no need for progress notifications, and if the tool list changes you refresh it in the client.
  • Two kinds of errors. Protocol errors, such as malformed JSON, mean the client has a bug. Tool errors, such as a team that does not exist or a record you cannot read, go back to the model so it can fix its call. That is what I saw in the tests: refused calls came back as normal responses marked isError, and Claude retried.
  • Multi-company. The initial context tool returns the user’s default company and the companies they can use. The agent works in the default one unless you name another.

What the agent can do by default

Out of the box, five tools are exposed, and all of them read:

  • ai_tool_mcp_retrieve_initial_context: who the user is, which companies, timezone;
  • ai_tool_get_models and ai_tool_get_fields: what data exists and how it is structured;
  • ai_tool_search: read records;
  • ai_tool_read_group: totals and counts, grouped.

Everything else, including creating and updating records, stays hidden until an admin ticks Available in MCP. The list of tools in the database is long: accounting report tools, menus, views, dates, website editing, image generation. Each one has to be switched on separately.

Adding your own tool

Because tools are server actions, you can add one without writing a module. The talk built a “create a helpdesk ticket” tool this way:

  1. Create a server action on the AI Tool model, of type Execute Code. Odoo names its own tools AI Tool: …, and the talk suggests the same.
  2. In the Usage tab, give the tool the name the agent will see, and a description. The description is the prompt the agent reads to decide when to call the tool, so say what it does, when to use it, what it returns and, if useful, when not to use it.
  3. Describe the parameters as a JSON schema (Edit lets you paste one).
  4. Write the code that runs when the tool is called.
  5. Tick Available in MCP.

Then let an agent try to break it. In the talk, Claude suggested edge cases for the new tool and found one at once: the tool accepted a helpdesk team that does not exist and created the ticket anyway. Odoo’s built-in rights still apply, but checking the inputs is up to the tool’s code.

Six real requests

For these runs I used Claude Code 2.1.319 (Claude Opus 5.5), connected with an MCP API key. An OAuth sign-in gives the same kind of key, with the same scope and the same tools, so the results would not differ. Two users:

  • Mitchell, the demo admin;
  • Sam, a new internal user with Sales: own documents only and nothing else.

Each request ran once, in Claude Code’s non-interactive mode (claude -p "<request>" --mcp-config <config> --allowedTools "mcp__odoo__*") so every tool call could be recorded. Claude Code normally asks before calling a tool you haven’t allowed; here the Odoo tools were allowed up front, and shell and file access were turned off. Tool calls below count only calls to Odoo. Times and costs are what Claude Code reported, at API prices.

Request User Tool calls (errors) Time Cost Outcome
1. Overdue invoices Admin 5 (0) 42 s $0.28 Full list, spotted credit notes
2. Top salespeople Admin 7 (0) 20 s $0.19 Ranking with names
3. Customer call brief Admin 28 (6) 84 s $0.71 Useful brief, many retries
4. Create a quotation Sam 9 (1) 21 s $0.20 Draft S00022 created
5. Everyone's sales and settings Sam 4 (1) 22 s $0.15 Correctly refused
6. Change a due date Admin 3 (0) 18 s $0.20 No update tool, explained how

Most simple questions took about 20 seconds and 20 cents. The open-ended customer brief took four times as long and cost more than three times as much. The answers below are shortened; the numbers are Odoo’s demo data.

1. “Which customer invoices are overdue?”

Which customer invoices are overdue today? Give me a short table: invoice, customer, due date, amount still due, days late. Then tell me which customer to call first.

The agent asked for the user’s context, then searched posted, unpaid customer invoices due before today. Without being asked, it also looked for open credit notes and found that some invoices were matched by credit notes of the same amount.

14 invoices are overdue as of 2026-11-09, totalling $218,571.14 across both companies. However, $112,582.51 of that (marked †) is probably already offset by open credit notes, leaving about $105,988.63 to collect.

Invoice Company Customer Due date Still due Days late
INV/2026/00004 † YourCompany Acme Corporation 2026-10-22 36,512.50 18
INV/2026/00005 YourCompany OpenWood 2026-10-31 2,000.00 9
INV/2026/00010 My US Company Azure Interior 2026-11-01 4,142.50 8
INV/2026/00002 YourCompany Acme Corporation 2026-11-04 48,012.50 5
INV/2026/00002 My US Company Acme Corporation 2026-11-04 44,150.63 5

(5 of 14 rows shown.) † An open Acme credit note matches this invoice’s exact amount.

Call Acme Corporation first. Acme owes $92,163.13 after those credits, all on INV/2026/00002. Before calling, have accounting confirm the credit notes do apply, so you don’t chase invoices that were already credited.

This is the use case from the first post, and it worked well: one question, five calls, and an answer an accountant can act on. The agent matched credit notes by amount only, and said so.

2. “Who sold the most this year?”

Who sold the most this year? Rank our salespeople by total confirmed sales orders, with the number of orders and the amount.

The agent grouped confirmed orders by salesperson. Odoo returned the salespeople as internal IDs (res.users(2,) and res.users(5,)), so the agent made a second call to look up their names before answering. The other calls read the user context and the order fields, looked up the currency, and checked the second company and the yearly total:

Marc Demo sold the most in 2026, with $10,027.00 in confirmed sales orders, about 55% of the total.

Rank Salesperson Orders Total (incl. tax)
1 Marc Demo 11 $10,027.00
2 Mitchell Admin 4 $8,198.13

Correct, at the cost of one extra round trip. A tool that returned names would save a call on every grouped question.

3. “Brief me before a customer call”

I have a call with Azure Interior in 10 minutes. Give me a one-screen brief: their open quotations and orders, unpaid invoices, and anything I should be careful about.

This was the heaviest run: 28 tool calls in 84 seconds. The agent checked the customer, contacts, orders, invoices, invoice lines, payments, activities and chatter messages. Six calls failed: two asked for fields that do not exist in this version, one used a malformed filter, and three asked for apps (CRM, Helpdesk, Inventory) that are not installed in the test database. The agent retried or moved on each time.

They effectively owe $4,142.50, and even that is debatable. Don’t chase the larger invoices until Accounting has looked at them.

  • Both INV/2026/00001 invoices (one per company) are exactly matched by unapplied credit notes.
  • Possible double billing: both companies invoiced the same 5 Large Meeting Tables and 5 Four Person Desks.
  • INV/2026/00010 has no payment terms, so check the agreed terms before calling it late.
  • Not checked: CRM, helpdesk and delivery records aren’t accessible from this Odoo connection.

The brief was useful, and its last line said what the agent could not see instead of pretending the picture was complete. Open-ended questions over many kinds of records take more time and money.

4. “Create a draft quotation”

With Create Records exposed, Sam asked:

Create a new draft quotation for the customer Azure Interior with 3 units of the product ‘Office Chair Black’ (or the closest office chair you find).

The agent found the customer and the product (its first product search used true instead of Python’s True and was refused, so it retried), created the quotation, added the line in a second call, and read the result back to check it. Each create call carried an explanation written for a human, such as “Create a new draft quotation for customer Azure Interior (partner ID 16), to which 3 units of Office Chair Black will be added. Do you confirm?”. In normal use, that is the text your AI app shows you before you approve.

The quotation number is S00022. It’s saved as a draft for Azure Interior. [FURN_0269] Office Chair Black, 3 × $120.50 = $361.50, total $415.73 including tax.

In Odoo, the quotation is there, created by Sam:

Odoo quotation S00022 for Azure Interior with 3 Office Chair Black at 120.50, total 415.73; the chatter shows Sam Sales (MCP test) created the sales order; margin shows -49%
S00022, created by the agent for Sam. Nothing marks it as made by AI, and the margin is -49%.

The chatter says Sam created it, with nothing to show it came from an AI agent. And the margin is -49%: in the demo data the chair costs more than its sale price. The agent did what it was told and did not mention it, so a human still has to review the quotation before sending it. The AI checks that it can do something, not whether it makes business sense. That is the best reason to keep write tools off, or to give them only to people who check what the agent made.

5. “Show me everyone’s sales”

Show me the total sales of every salesperson in the company, and list the system parameters of this database.

I couldn’t get either result: this Odoo connection runs as Sam Sales (MCP test), and that account can’t see company-wide sales or the system parameters. There are no confirmed sales orders visible to this account. It sees only two draft quotations, both its own. […] Odoo returned an error saying ir.config_parameter “doesn’t exist or is inaccessible to the current user”. […] I didn’t try to reach it another way.

The one error in the table is Odoo refusing the system parameters, which is the expected answer. The record rule (own documents only) and the access rights did their job, and the agent explained the limits instead of presenting two quotations as “the company’s sales”. This is the result I most wanted to see.

I also checked the edges directly against the server. Sam could not create a contact: that is Odoo’s access rights on contacts, not the tool’s Allowed Groups (Odoo answered with the groups that would allow it), and no one, admin included, can create system parameters through MCP: Odoo’s AI tools refuse all ir.* technical models (except reading menus and attachments) and a few others such as user groups.

6. “Change the due date of this invoice”

Change the due date of invoice INV/2026/00008 (YourCompany) to 2026-11-20.

Update Records was not exposed (only the five reading tools and Create Records were), so the agent could not do it:

I couldn’t change the due date, so INV/2026/00008 still shows 2026-11-06. The Odoo connection here can only search and create records, and has no way to edit an existing one. I didn’t create anything as a workaround, because a new or duplicate accounting record wouldn’t change this invoice.

It then gave the steps to do it by hand and warned against resetting a partially paid invoice to draft. The exposed tool list, not the model, decided what the agent could do.

What it still lacks

The tests confirmed the most important part: the AI always works with the signed-in person’s rights, never as a super admin, and every result followed Odoo’s access rights and record rules. The data tools also spent no IAP credits, because the language model is the one in your AI app; Odoo only runs the tool. Several controls that the community modules already had are missing, though.

Read-only AI access exists only through groups. Once an admin turns on Create Records, everyone who can use it creates whatever their Odoo rights allow; people outside its Allowed Groups get the reading tools only. What is missing: there is no read-only choice when signing in, no limit per app or per token, and no limit per kind of record. You also cannot let someone write in Odoo by hand but only read through AI, unless you create a group for it. That is the workaround: create an empty “AI writers” group, put it in Allowed Groups on each write tool, and add only the people who should write through AI.

Odoo does not ask before writing. Whether anyone approves a change depends on the AI app and how it is set up, as in request 4.

Odoo’s team confirmed these points in the talk’s Q&A. A tool cannot be turned on for some users only: once it is on, it is on for everyone (they did not mention Allowed Groups, which limits it by group, as tested above). Confirmation is the AI app’s job, and the server runs whatever it is asked to. There is no throttling of MCP calls.

Odoo cannot show you later what the AI did. A quotation made by an agent looks exactly like one Sam typed, and there is no screen listing agent activity.

It needs Odoo 20 Enterprise for OAuth. Odoo Online 19.4 has the server with API keys only, and 17.0, 18.0, 19.0 and Community do not have it.

For developers, the details behind those points:

  • Exposure is the use_in_mcp flag on ir.actions.server; the server action’s group_ids are checked both for tools/list and tools/call (tested: the tool disappears from the list, and a direct call returns isError with an access-rights message).

  • Write tools are marked destructiveHint for the client, and the dispatcher calls them with tool_request_confirmed=True, which skips Odoo’s own confirmation step. Tools run with env(su=False).

  • The only audit trail is one server log line per call, with no user and no arguments:

    INFO mcp20 odoo.addons.ai_mcp.controllers.mcp_controller: tools/call:ai_tool_get_models 0.046s success 127.0.0.1
    
  • There is no rate limit in ai_mcp.

  • Filters must be written as Python literals (True, not true), grouped results return record references instead of names, and error messages include the internal Python expression that failed. Claude worked around each of these, but every retry costs time and tokens.

Should you turn it on?

Yes for reading, on Odoo 20 Enterprise. It comes with the AI app, OAuth works without setup for the main AI apps, and every answer stayed within the signed-in person’s rights. Leave the write tools off until you know which ones you need, limit them with Allowed Groups, and connect agents through users with narrow rights instead of relying on the AI app to ask first. Request 4 is the reason: the agent created a quotation at a loss and did not say so.

If you need per-model limits, a read-only grant or an audit screen, or you run Community or an older version, the community modules still do more. Connecting AI Agents to Odoo with MCP, Seven Months Later compares them with the official server and says what we would pick in each case.

  1. MCP server (20.0 documentation) — Odoo ↩︎

  2. MCP tools (20.0 documentation) — Odoo

  3. MCP server (saas-19.4 documentation) — Odoo

  4. Authorization (2025-11-25 specification) — Model Context Protocol

  5. Claude Code client metadata document — Anthropic

  6. Odoo MCP: A technical deep-dive into AI connectors (YouTube, 30 September 2026) — Odoo ↩︎

Ready to put AI to work?

Let's explore how Trobz AI can automate your processes, enhance your ERP, and help your team make better decisions — faster.