Odoo 20 MCP Server: Connect Claude or ChatGPT to Odoo Safely

Most Odoo MCP articles stop at "AI can now talk to your ERP". This guide covers the exact setup for Claude and ChatGPT, the default tools, how rights and approvals really work, and when a community server fits better. Our AI and Odoo integration services deliver all of it.

It is written for IT managers, Odoo administrators and founders who already pay for an AI assistant and want it working on live Odoo data. For how Odoo 20 bills its own AI, see our Odoo 20 AI features and IAP credits breakdown first.

What is the Odoo 20 MCP server?

The Odoo 20 MCP server turns your database into a remote Model Context Protocol server that an external AI agent can call through a standard interface. You keep working in Claude, ChatGPT or Codex; the agent calls Odoo tools behind the scenes.

MCP (Model Context Protocol) is an open standard for connecting language models to applications. In Odoo's AI MCP server documentation, the database is the server and your AI app is the client, loading the exposed tools once per session.

The tools are ordinary server actions with a flag set; our Odoo 20 AI release notes record the /mcp route and the mcp key scope. Unlike Odoo's built-in agents, the model runs in your own tool, apart from the classic external API covered in our Odoo API integration guide.

Which Odoo versions and plans include the MCP server?

The native MCP server exists in Odoo 20.0 and in Odoo Online from saas-19.4, on the Enterprise Custom plan. Everything older, and every Community database, needs a third-party server instead.

Your setup	Native /mcp endpoint?	What to do
Odoo 20 Enterprise, any hosting	Yes, with the AI app installed	Follow the setup below
Odoo Online saas-19.4	Yes, documented	Same setup; check your version at odoo.com/my/databases
Odoo Online 19.0 to saas-19.3	No, /mcp returns 404	Wait for the SaaS upgrade or use an external server
Odoo 19.0 on Odoo.sh or on-premise	No	Community module, or upgrade to Odoo 20
Odoo 16 to 18, any hosting	No	Community module or external server over the API
Odoo Community, any version	No	External server over the API

Odoo publishes the MCP server page for 20.0 and saas-19.4 only; the 19.0 to 19.3 documentation has no such page. One Odoo Online user had the MCP key scope but found the /mcp endpoint returned 404 on an older SaaS build.

Plan matters too. Odoo's AI overview warns that installing the AI app on a Standard subscription triggers an upsell to Custom, and a forum answer confirms that Odoo 20 MCP requires the Custom plan on Online, Odoo.sh and on-premise.

In the US that is US$31.10 versus US$61.00 per user a month billed yearly; our Odoo pricing guide has the full comparison.

How do you connect Claude or ChatGPT to Odoo 20 with MCP?

You connect Claude or ChatGPT to Odoo by pointing the client at your database URL plus /mcp and authenticating with OAuth or an MCP-scoped API key. Odoo documents both routes and recommends OAuth.

Method	Who can use it	Needs Node.js?	Best for
OAuth from the client	Internal users only	No	Claude Code, ChatGPT, Grok
API key with MCP scope	Internal and portal users	Yes (npx mcp-remote)	Claude Desktop, Codex, Antigravity

The OAuth route, step by step:

1. Approve the client. In Settings, Integrations, open MCP OAuth and add the client's metadata URL. Leave Allow Dynamic Client Registration off, because it lets any client register.
2. Add the server in the client. In Claude Code run `claude mcp add --transport http odoo https://yourdb.odoo.com/mcp`. In ChatGPT, go to Settings, Plugins, Add custom MCP server, enter the URL and choose OAuth.
3. Grant access. The client opens the Odoo login page. Sign in, choose how long access stays valid, and click Allow.
4. Check the connection. Run `claude mcp list` or open the client's MCP panel; the Odoo server should show as connected.

The API key route: open My Preferences, Security, Create API Key, choose the MCP scope and an expiry, and copy the key once. Then paste Odoo's mcp-remote snippet into claude_desktop_config.json, or add it in Claude Code with `claude mcp add --transport stdio` and `npx -y mcp-remote`.

What can an AI agent do through Odoo MCP tools?

By default an agent can list models, read fields, search records and run grouped reports; it can create or update records only after an administrator exposes those tools. The available MCP tools page lists 27 tools.

Tool group	Examples	Exposed by default?	Writes data?
Setup and context	MCP Retrieve initial context (user, timezone, companies)	Yes	No
Data querying	Get Models, Get Fields, Search, Read group	Yes	No
Data querying (write)	Create Records, Update Records	No	Yes
Navigation	Open Menu List, Kanban, Graph, Pivot, Compute Date	No	No
Website styling	Get Snippets, Write Custom CSS, Apply HTML to Page	No	Two of four
Media and extras	Generate an Image, Search Images, Web Search, Load Skills	No	Image generation only

Every other tool is hidden until you open Settings, Technical, Server Actions in developer mode, pick it and tick Available in MCP on its Usage tab. There is no delete tool and no "run any method" tool in the native list, which for most companies is a feature.

How do access rights and human approval work with Odoo MCP?

Every MCP tool call runs as the Odoo user who authenticated, and the approval prompt before a write comes from your AI client, not from Odoo. That split is the whole security model.

• Identity: each person connects as themselves. Access rights, record rules and field restrictions apply exactly as in the web client.
• Expiry: OAuth grants and API keys both expire. The Odoo 20 code offers 1 day, 1 month or 1 year, with a "Never" option visible only to system administrators.
• Revocation: every connection appears as an API key under My Preferences, Security. Delete the key and the client is cut off.
• Readonly Tool flag: it tells the client the tool is safe to run without asking. Odoo's documentation is explicit that it does not hide the tool; it is advice, not a lock.
• Client approval: Claude Code, Claude Desktop and Codex ask before running a tool not marked read-only. That prompt is your human gate for Create Records and Update Records.
• Audit: the 20.0 documentation describes no MCP call log. Keep the client's tool log and add logging to any custom write tool.

Our rule: never tick Readonly Tool on anything that writes, and grant writes through narrow custom tools. One forum contributor flagged a risk that is easy to miss: a write that triggers a customer notification sends an email nobody reviewed unless the connector previews it.

How do you expose a safe custom write tool?

A safe write tool is a server action that creates a draft or schedules an activity, declares its arguments in the AI Schema, and leaves posting to a person in Odoo. Odoo's AI server actions guide explains the pattern.

The AI picks the tool and the arguments; the tool's Python code does the work. Odoo states that a called tool "will execute unconditionally, unless the code itself prevents it", so business rules belong in the tool. Arguments must be declared on the Usage tab under names that match the code.

An accounts payable answer on the Odoo forum shows the shape: instead of exposing account.move, define narrow tools such as create_vendor_bill that create drafts and run existing validation actions under a dedicated integration user.

Our checklist for client tools:

1. One tool, one step: search purchase orders, create a draft bill, request a credit hold. No generic "write any record".
2. Drafts only: the tool never posts a journal entry, confirms an order or changes a price list; it schedules an activity for the owner.
3. Declared arguments: every parameter sits in the AI Schema with a clear description.
4. Logging: the tool writes a chatter note with the client, the user and the arguments.
5. Test with the real client: confirm the approval dialog in Claude or ChatGPT shows what will change.

Odoo 20 native MCP vs community Odoo MCP servers

The native server wins on simplicity and vendor support; community servers win on older versions, Community edition, per-model write control and audit logs. Pick by where your database sits today.

Criterion	Odoo 20 native MCP	Community MCP servers
Versions	Odoo 20.0 and Online saas-19.4	Odoo 16 to 20, vendor dependent
Edition and plan	Enterprise, Custom plan	Community and Enterprise
Where it runs	Inside Odoo at /mcp	Odoo module, or external process over the API
Default access	Read tools on, write tools off	Per-model read, create, update, delete toggles
Authentication	OAuth or MCP-scoped API key	API key, OAuth 2.1 or password, by product
Governance	Client log only; no documented call log	Audit log, rate limits, IP allowlists on some

The best known open-source bridge is mcp-server-odoo, an MPL-2.0 Python package run with uvx. In production it pairs with the free MCP Server module on the Odoo App Store, listed for Odoo 16.0 to 20.0 with per-model permissions, OAuth 2.1 and audit logging. Its "YOLO mode" bypasses every control.

Other options named by their vendors on the Odoo forum include Niyu Labs' MCP Server Connector, Safe Claude Connector and the hosted Odoo MCP Pro. We have not tested each one; compare versions, Odoo Online support, write controls and logging before you choose.

Two constraints decide most cases. Odoo Online cannot install custom modules, so Online users below saas-19.4 need an external server over the API. And XML-RPC is deprecated in Odoo 20, with removal planned for Odoo 22, so prefer JSON-2. Our Odoo integration services run that evaluation.

Does the Odoo MCP server use Odoo AI credits?

No, the core MCP tools run inside your database and the model runs in your own AI client, so you pay your Claude or ChatGPT provider, not Odoo IAP. That is the main financial reason to choose MCP over Odoo's built-in agents.

An Odoo forum answer on Odoo IAP versus external AI subscriptions puts it plainly: if you already pay OpenAI, Anthropic or Gemini, connect that subscription through MCP. Your data is then handled under your provider's agreement, not Odoo's.

One caveat: tools that call Odoo's own AI service, such as Generate an Image and Web Search, may draw Odoo AI credits. The documentation does not say, so test first.

Odoo MCP use cases by industry

The highest-value MCP use cases are read-heavy questions that today end in a CSV export, followed by narrow draft-creating actions. The table shows where we see the fastest payback; our practical Odoo AI use cases go deeper.

Industry	Pain or cost driver	Odoo approach	How iVentureTeam helps
Distribution	Daily stock and margin questions answered by exports	Search and Read group on stock and sales data	Prompt library and read-only key policy
Manufacturing	Planners checking MPS gaps by hand	Read group on forecasts, custom "flag shortage" activity tool	Custom server actions with declared arguments
Professional services	Managers chasing budget and timesheet status	Navigation tools open filtered pivots on demand	Agent skills per role, approval rules
Accounts payable	Bills keyed from PDFs, approvals by email	Draft-bill tool run by a dedicated integration user	Tool design, logging, UAT with finance
eCommerce and retail	Website edits queued for a developer	Website styling tools with explicit approval	Exposure review, staging-first rollout
Healthcare and SaaS support	Agents answering tickets without ERP context	Read tools for customer and contract data	Hybrid agent across helpdesk and Odoo

Common Odoo MCP setup problems and fixes

Most Odoo MCP failures come from the wrong version, the wrong key scope or a missing client prerequisite, not from the protocol. Check these before opening a ticket.

• /mcp returns 404: the database is below saas-19.4 or 20.0, or the AI app is not installed. Check the version at odoo.com/my/databases.
• Key accepted, tools missing: the key has the rpc scope, or the tool lacks Available in MCP. Scope is enforced, so create a new key with the MCP scope.
• mcp-remote fails to start: Node.js is missing on the machine running Claude Desktop or Codex.
• Agent sees too much: the connected user has broad rights. Create a dedicated user with limited groups instead of using an administrator.
• Writes happen without a prompt: a write tool was marked Readonly Tool. Untick it; the flag only advises the client.

Not sure which version your database runs? Our free Odoo health check tells you whether native MCP is an option yet.

How iVentureTeam helps with Odoo MCP server setup

iVentureTeam sets up the Odoo 20 MCP server, designs the tool and approval model, and builds the custom server actions your agents need. We have delivered 150+ ERP projects across 35+ countries with 40+ Odoo specialists over 13+ years.

• Readiness check: version, edition, plan and hosting reviewed against the native MCP requirements, with a community-server plan if you are not there yet.
• Secure rollout: OAuth client approval, dedicated agent users, key expiry policy and a documented list of exposed tools.
• Custom tools: narrow, draft-only server actions with declared arguments and logging, delivered through our AI and Odoo integration services.
• Agent design: prompts, skills and escalation rules for Claude or ChatGPT, built by the team behind our AI agent development services.

Odoo MCP server case study: LabMart

LabMart shows the human-in-the-loop pattern this guide recommends, applied to a support agent working across Zoho Desk and Odoo. The healthcare SaaS company operates in the USA and India.

Tickets had grown four times year on year and first response had slipped from 2 hours to 18 hours. We built a multi-LLM classifier, including Claude Sonnet 4 for clinical-language edge cases, wired into Zoho Desk and Odoo Helpdesk.

The controls matter as much as the model. Anything below a 78% confidence score is flagged for human review, specimen-safety tickets bypass automation entirely, and a live audit dashboard logs every AI decision.

The agent went live in 7 weeks. It handles 1,800 tickets a week, auto-classifies 92% of them and cut first response from 18 hours to 4.5 minutes. Read the full LabMart AI agent and Odoo Helpdesk case study.

The bottom line on the Odoo MCP server

If you run Odoo 20 Enterprise on the Custom plan, switch on the native MCP server today for read-only use and add write tools one narrow action at a time. On Odoo 19 or Community, a community server gets you the same workflow now, with more governance features.

Either way the rules hold: each person connects as themselves, writes create drafts, approvals happen in the client, and someone owns the log. Done that way, MCP is the lowest-cost AI project you can run on Odoo.

Ready to connect your AI assistant to Odoo without the risk?

Get a tool-by-tool MCP plan for your Odoo database. Request an Odoo MCP readiness review, call +91-90235-15208, or email business@iventureteam.com.