Odoo Updates · Odoo 20

Odoo 20 AI agents and the MCP server: what they do and what they cost

A deep dive for owners, IT leads and anyone asked to "add AI" to Odoo: what an agent can do on your records in version 20, how agents run inside automations, how outside AI tools connect through MCP, what it all costs, and how to keep it under control.

By Octagotech Team10 min readUpdated

Key takeaways

  • Odoo 20 agents can create and update records, take a file as instructions and answer questions about a document you are previewing.
  • Automated and scheduled actions can now call an agent, so AI becomes a step in a process that runs by itself.
  • The MCP server lets tools such as Claude or Codex work in your database with an MCP-scoped API key, through the tools you expose.
  • Every AI feature now needs IAP credits. Off Odoo Online, confirm with Odoo whether you also need your own OpenAI or Gemini key.
  • Treat each agent like a new user: narrow skills, a test database, and a written list of what data leaves.

Our Odoo 20 overview gave AI two paragraphs. This deep dive covers the rest: what Odoo 20's agents actually do, how they plug into automations, what the MCP server exposes, what you pay, and the controls we put around them. The facts come from Odoo's Odoo 20 release notes, the AI section of the 20.0 documentation, Odoo's launch post and the 20.0 source code, checked on September 30, 2026.

What an AI agent is in Odoo 20

Odoo's AI agents documentation describes an agent as a configured assistant with three parts. A system prompt sets its role and tone. Skills, which version 19 called topics (the release notes record the rename), bundle instructions with tools, the functions that actually do things in the database. Sources are the material it reads from: PDFs, web links, files in Documents and Knowledge articles. An option called Restrict to Sources keeps its answers inside that material.

Odoo ships skills for building views, retrieving information, creating records, updating records, generating images, searching the web, editing website pages and, with CRM installed, creating leads. The documentation is explicit that an agent with no skills can only give information; it cannot change anything. Creating or customizing agents needs the AI app installed, and that app is not part of the Community source for 20.0, so plan on Enterprise.

What is new in 20 is mostly what agents can do on records. The release notes list creating records on request, updating records, filtering by weeks or quarters when building a view, and generating images and enriched buttons for websites and mailings. Agents also ask questions and permission through clickable buttons, stop for confirmation at their tool call limit, and accept dictated requests. Conversations are kept for up to 30 days after the last message.

Files as instructions, and questions about a previewed document

Two changes turn documents into inputs. First, you can ask an agent to create records by uploading a file of instructions, such as a PDF. Second, you can open a document in preview and ask questions about it. Per the documentation, a file attached to a conversation counts as a source for that conversation only, unless you add it to the agent's sources. A previewed file in Documents becomes context for the chat, combined with the agent's own sources.

Who this suits: anyone who re-keys the same kind of document every week, such as a supplier price list, a customer's order sheet or a project brief. What to check: how the agent handles a messy scan, a second page, or a field it cannot find. And when a source document changes, reprocess it, which the release notes now let you do by hand, or the agent answers from the old version.

Agents inside automated and scheduled actions

This is the change with the most reach. The release notes say agents can now be called by automated and scheduled actions, that you can review the automation chats to see how an agent handled each run, and that you can ask for changes based on the errors it hit. Odoo's launch post, Meet Odoo 20, describes the result: you tell an agent in plain language what process to run, it explains what it is about to do, and the process then runs on its own. Its examples include confirming a sales order when the matching email arrives and routing leads to the right salesperson.

Underneath, the 20.0 documentation on AI server actions sets out a manager and worker split. An AI server action reads the record, interprets its prompt, and chooses which tool to call and with what arguments. The tools are ordinary server actions marked for use in AI, and they do the writing. Two sentences in that page matter more than any feature list: a tool called by the AI runs unconditionally unless its own code prevents it, and business rules must be enforced in the tool's Python code. The AI chooses; your code decides what is allowed.

The release notes also say agents can now "update themselves" to serve their use case better. Neither the notes nor the 20.0 documentation explain what an agent changes, whether that is its prompt, skills or sources, or whether a person approves the change. Until Odoo documents it, we treat it as a configuration change to review: record an agent's settings before a run on a test database, compare them afterwards, and read the automation chats.

An automation with AI in it runs, and spends, on every trigger. A rule that calls an agent on each new lead or ticket uses credits every time it fires. AI fields add a quieter cost: the documentation says a scheduled action fills empty AI fields once a day and is active by default. Estimate triggers per month before you switch anything on.

AI features inside specific apps

Beyond the AI app itself, the release notes place AI in these apps:

  • Documents: rename uploaded files with AI, on top of the AI document sorting already described in the documentation.
  • eCommerce: ask the assistant to generate product images or change a product's name, price or description from the Shop page.
  • Helpdesk: spot similar tickets and suggest answers from them.
  • Marketing Automation and Social Marketing: help setting up campaigns, and drafting social posts.
  • ESG: assign the most suitable emission factor to each emission.
  • Website: the AI Website Assistant builds pages, forms and menus from a prompt, marks the blocks it made, and helps with SEO; our retail, POS and eCommerce deep dive covers it.
  • Live Chat: when an agent mentions a record, the chat can show a clickable preview card.

Odoo's launch post also describes an accounting agent that answers questions from your reports, such as receivables and cash flow, and a purchasing agent that reviews vendor history. Our accounting deep dive and operations deep dive cover those areas. Note that bill line prediction and suggested reorder levels are described in the notes as suggestions from your history, not as AI features.

The MCP server: letting outside AI tools in

MCP, the Model Context Protocol, is an open standard for connecting AI tools to other software. In Odoo 20, your database can act as an MCP server. The MCP server documentation walks through the setup: a user creates an API key with the MCP scope and a validity period in their preferences, then configures the outside client with that key and the database address followed by /mcp. The documented clients are Claude (Desktop and Claude Code), Google Antigravity and OpenAI Codex, and the client machine needs Node.js.

What the client can do depends on what you expose. By default, only a handful of read tools are visible: listing models and fields, searching, grouping, and loading the user's starting context. Everything else, including creating and updating records, driving the interface and editing website styles, stays hidden until someone ticks "Available in MCP" on that server action. A second box, Readonly Tool, only tells the client a tool is safe to run without asking; it hides nothing. The key authenticates the user's identity and permissions, and Odoo's launch post says outside tools work "with all your access rights intact".

Two practical limits. Odoo's external API documentation says no API key can last more than three months, so plan rotation. And that page says external API access needs a Custom plan; the MCP page does not say which plan it needs, so confirm with Odoo before building on it from a Standard subscription.

What AI costs in Odoo 20

The release notes are short on this: all AI features now need credits, bought through In-App Purchases (IAP). Odoo's Odoo AI listing sells credits in packs, starts a database with 20 free credits, and says usage depends on the feature, how complex the task is and how large the request is. It also offers low-balance warnings and automatic refills, both worth setting up before any automation goes live.

Hosting changes the picture. According to the 19.0 documentation, Odoo.sh and on-premise databases can configure their own OpenAI or Google Gemini API key for AI features, billed by the provider. The 20.0 version of that page was still being written on September 30, when it pointed to the Odoo AI credits listing. Until Odoo settles it, ask whether a version 20 database off Odoo Online pays in credits, in provider fees or both. Self-hosted databases also need the pgvector PostgreSQL extension for AI features, according to Odoo's 20.0 installation guide.

Then the plan. On September 30, Odoo's pricing page listed "Agentic AI" next to Studio as a Custom plan feature, and its FAQ ties the Custom plan to Studio, agentic AI, custom development and the API. The AI fields documentation adds that installing Studio may change your plan.

AI capability, where it lives, what it needs

AI capabilityWhere it livesWhat it needs
Ask AI and agent chats: answer, create and update records, read filesCommand palette and AI button across apps; Documents previewIAP credits; Create or Update Records skill for changes
Custom agents with your prompts, skills and sourcesAI appAI app (Enterprise); credits
Agents called by automated and scheduled actionsAutomation rules, scheduled actions, AI server actionsCredits on every run; tools marked for AI, with checks in code
AI fieldsStudio or property fields on a recordCredits, including a daily fill of empty fields; Studio may change your plan
Document sorting and file renamingDocumentsDocuments app; credits; the AI app is not required
Live chat agentLive Chat channelsLive Chat and AI apps; credits
Website Assistant and image generationWebsite editor, Shop page, mailingsCredits
App features: similar tickets, campaigns, social posts, emission factorsHelpdesk, Marketing Automation, Social Marketing, ESGThe app; credits
MCP server for outside AI clientsYour database address plus /mcpMCP-scoped API key (three months at most); exposed tools; confirm the plan with Odoo
Any AI feature off Odoo OnlineOdoo.sh or your own serverPer the 19.0 docs, your own OpenAI or Gemini key; pgvector on self-hosted PostgreSQL

Governance: access, testing and data

Access. Give each agent the fewest skills its job needs, and keep write tools out of MCP until there is a named reason to expose them. Create MCP keys per person or per integration, never shared, with short validity, and revoke them when someone leaves. Odoo 20 also simplified access rights, with record rules replaced by a domain on each access right, so re-test what each role can see after the upgrade; our developer changes deep dive explains that change.

Testing. Build on a copy of your database and run each use case against real records with the people who will use it. Compare the agent's output with what that person would have done, then try the awkward cases: a missing field, two possible matches, a document in another language. Record the credits one run uses and multiply by volume. Keep a person approving anything that posts, pays or emails a customer until the error rate is known.

Data leaving the database. When an AI feature runs, the prompt and the record content it needs go to the AI provider behind it; Odoo's 19.0 documentation names OpenAI and Google Gemini. With MCP, results also go to the outside client and whichever model runs it, so your data policy has to cover that tool too. Some limits are documented: voice recordings are discarded once transcribed, and agent conversations are kept for 30 days. Write down, per use case, which records are sent and where, and have the data owner approve that list.

Where AI pays off first, by industry

AI earns its keep where there is repeat work and enough history in Odoo to learn from. That is the brief of our AI strategy service, and version 20 widens the options in each sector it names. In manufacturing, the first candidates are summaries of quality notes and demand forecasts that inform reorder rules. In distribution, it is vendor bill and document capture at volume, now with agents inside the automations that route those documents. In multi-location retail, it is an agent for product and order questions, plus AI help with eCommerce product pages. For service teams, the natural starting points are Helpdesk's similar-ticket suggestions and a live chat agent restricted to your own articles.

In each case clean data comes first, because an agent reads what is in Odoo. We have no published AI case study yet, so we scope each use case, test it on your records and let it go live only if it holds up.

Should you wait for Odoo 20 to start?

Usually not. AI agents and AI server actions arrived in Odoo 19, so a 19 database can start on sources, prompts and a first use case now. Our advice on version timing is unchanged: land on 19 first and move to 20 once the first rounds of fixes are in, unless a version 20 feature, such as agents inside automations or the MCP server, is critical to you. Then test it on an upgraded copy and run it beside live first. Odoo Online databases on a 19.x online version will move to 20 on Odoo's schedule, so start testing early. Custom tools your agents call must be ready for 20 too; our Odoo customization team plans them alongside the Odoo upgrade.

Frequently asked questions

Do Odoo 20 AI features cost extra?

Yes. Odoo's release notes say every AI feature now needs credits bought through In-App Purchases, and usage varies with the feature and the size of the request. On Odoo.sh and on-premise you can configure your own OpenAI or Gemini key, according to the 19.0 documentation, and the 20.0 page was still being written on September 30, so confirm version 20 billing with Odoo before you budget.

What is the Odoo MCP server?

It lets an outside AI client, such as Claude, Google Antigravity or OpenAI Codex, read and write your Odoo data. The client connects to your database address plus /mcp with an MCP-scoped API key, and it can only call the tools, which are server actions, that you expose to it.

Can Odoo 20 AI agents change records on their own?

Yes, if you let them. An agent with the Create Records or Update Records skill can change data, and agents can now be called from automated and scheduled actions. An agent with no skills can only answer questions, so give each agent only the skills its job needs.

Do we need Odoo 20 to use AI in Odoo?

No. AI agents and AI server actions arrived in Odoo 19. Version 20 adds agents inside automations, the MCP server, files as instructions and more AI in individual apps. Land on 19 first unless one of those is critical to you, and test it on an upgraded copy.

Octagotech Team

Certified Odoo Ready partner · San Francisco, New York, Ahmedabad and Rajkot

Our AI strategy work starts with the data already in your Odoo. Our engineering team in Ahmedabad builds and tests each use case on a copy of your database, and a named project manager in New York scopes it with you. About Octagotech, Odoo Ready partner.

Weighing AI in Odoo 20?

30 minutes: bring one repetitive task and your Odoo version, and we will tell you whether AI is worth it there.

Book a discovery call

Find out where AI would actually save your team time

The first call takes 30 minutes.

Or email sales@octagotech.com