HELP WANTED: Odoo LLM agents are here!

[ayushin]

Hi all,

We're getting amazing results by connecting reasoning/agent models (ChatGPT, Grok, DeepSeek, etc.) to Odoo and enabling the models to interact with the Odoo database via functions for retrieval and existing server actions for database manipulation.

Some base modules—e.g., for model/provider configuration and streaming chat—are mature, while others are being prototyped and built as we speak.

https://github.com/apexive/odoo-llm

This approach has the potential to be a game changer for managing and customizing Odoo.

Please reach out if you’re interested in helping or testing. We intend to keep this core open source.

Alex | Apexive

[rvalyi]

awesome! Indeed the potential of getting an LLM querying Odoo is certainly huge. Great you reached the OCA for this... May be we should create an OCA repo for this? May be it's still a bit early?

@ayushin are you already familiar with contributing to the OCA? Did you or your company contributed some module already? some PRs you got merged may be?
Meanwhile you might want to read https://github.com/OCA/odoo-community.org/blob/master/website%2FContribution%2FCONTRIBUTING.rst

[ayushin]

i am happy to transfer the whole https://github.com/apexive/odoo-llm -> https://github.com/OCA/llm

But we need help with creating the unit tests / translations and so on - that's exactly the man power we don't have to jump through the hoops and hoping to find help here

[afsuyadi]

Hey there. I have seen big potential here, where LLM are integrated into Odoo environment. We could build lots of AI Agents which are very helpful. I am wondering if I could provide some help for this community, as I am now working as an Odoo Developer and I have just also started building RAG using LangChain. Looking forward to hear more. @ayushin

[kbb01]

Hi I've also starting using LangChain with RAG and wondered if Model Context Protocol could play a part in AI understanding the tools available for integrating into Odoo

[ayushin]

if you look at how the modules / models structure is setup (and feedback and PRs are welcome) - you will see we are building a foundation for an ecosystem that can make Odoo truly first AI-driven ERP

[rvalyi]

yeah from what I saw it doesn't look bad at all... Let's see what the other OCA contributors say...

cc @etobella @pedrobaeza @sbidoul @simahawk @alexis-via @sebastienbeau @vdewulf

[kbb01]

Hi happy to help, big potential here I feel.

[rvalyi]

@ayushin a repo/module that is worth checking and you could get your LLM agent interact with: https://odoo-community.org/shop/automation-oca-715899#attr=940733

[zamberjo]

Hi, first of all thank you very much for sharing it.

It's funny because lately I was studying how to apply this topic in Odoo, and I had started a side project, although I didn't integrate it in an Odoo module.

The issue of using APIs, IMHO, with production data is dangerous, because nobody assures you that there will be a leak of such data, they can train the next model using what was sent. For this reason I had started a project that runs locally, for lack of resources I had to use the “mistralai/Mistral-7B-Instruct-v0.3” model.

My main idea was to create a service with an API in FastAPI which uses an agent initialized with the ir.action.window.server as Tools. This way all data used as contex is in a sand box.

In addition IMO it would be necessary to control the blocking of all use of model.unlink() .

In general it is a good starting point to be able to integrate LLMs to Odoo. A repository could be created and everything could be discussed there.

Good work.

[ValentinVinagre]

Hello,
Very good work and beginning of LLM instruction in Odoo.
There are several points to consider:

• As the rest of the colleagues indicate, security is the main "Achilles heel". It would be a good point to address.
• Information roles.
• Contexts.
• Multiple models and agents.
• Reinforcement
• I see that the "pydanticAI" framework has not been used. Why? It could be a good starting link to normalize certain things.
• As I see it, it would also allow the connection with models created as proprietary, which is great.
• Expansion of models as needed, it would be necessary to see how it extracts the information and passes it to the model, etc.
• http calls

In general, I see it as a fairly well-oriented idea, great work. I will be happy to review the work carried out at OCA 😄

[rvalyi]

Hello, Very good work and beginning of LLM instruction in Odoo. There are several points to consider:

[...]

* Reinforcement

Would be great

* I see that the "pydanticAI" framework has not been used. Why? It could be a good starting link to normalize certain things.

Seems good indeed. Very trustable team also.

* As I see it, it would also allow the connection with models created as proprietary, which is great.
* Expansion of models as needed, it would be necessary to see how it extracts the information and passes it to the model, etc.

* http calls

for this It might also be worth considering the advanced PRs to get a better webhook system in Odoo (in Odoo CE webhooks start at v17 and we could have better ones in the OCA) https://github.com/OCA/webhook/pulls
These webhooks are both incoming and outgoing. So an Odoo AI agent could call user defined outgoing webhooks eventually... And/or eventually receive an incoming callback and trigger some AI action...

[ValentinVinagre]

I didn't have that repository under control, we'll keep an eye on it, thanks.
It could be a way to give it connections, although it depends on the architecture of the modules, both internal and external instructions could be given. On the other hand, it could be complemented with a telegram or whatsapp chat, which @etobella was doing, but is it interesting that it is integrated or would it be better to do it separately? All this will have to be assessed.
I think that before going off on a tangent, we would have to analyze and structure how it will impact the base system and give the possibility of creating local models or if finally, it would be perfect to create a working group on the OCA days, etc.
The whole world of AI is evolving very quickly (without taking into account security and many other important things) and it is not really known how it will end, since governments are also involved. I think it is perfect to have a generic module structure in OCA that gives versatility to use. However, you will have to be careful with everything you bring to the community, since they can be "dangerous" things in terms of data security.

[HekkiMelody]

I'm saddened to see all the uncritical hype for the "potential" of the lying, power-and-water-hungry plagiarism machines.

I was going to list all the ways in which LLM use would go against what I expect the values of an open source project to be (you know, fairness, data privacy, openness - since none of the LLMs are actually open source) but then I realized I couldn't find any "values" on the OCA website, so fair enough.

If you don't care for any of the moral arguments, LLMs WILL create technical debt for you, you WILL have to answer to your customers when they hallucinate, or when the action they "decide" to perform is wrong. LLMs are probabilistic, not deterministic, so you're just gambling.

[rvalyi]

@aleuffre : it's a bit like any tool: you can use nuclear power to build power plants and avoid burning coal or you can make bombs with it. As for LLMs: you can have stupid expectations or you might instead automate basic things effectively. I bet in less than 2 years nobody will want to choose an ERP where bank reconciliation will be manual or where no AI agent will help you query statistics in your database.

As for resources usages, models like Mistral or Deepseek use many orders of magnitudes less resources (both for training and for running the model) than the crazy amount of resources models like OpenAI were using just a few months ago. It can even run on a Rasberry PI for smaller models and I doubt we would need to train it ourselves over the whole Internet again just for using it in Odoo, so it's probably not more resource usage than running MRP or a website on Odoo v8...

Finally LLMs will be integrated with ERPs whether we like it or not I'm afraid. It can be done only in SAP and Odoo Enterprise and the OCA can marginalized or it can be done in the OCA in the best possible ways...

And as for a hyped technology that created a speculative bubble, Odoo itself is not bad either :-)

[Tardo]

I would be interested in integrating https://localai.io/ + https://docs.openwebui.com/ ... It may be interesting to explore the power of RAG's.

[ayushin]

As I expected, this discussion is expanding rapidly, and I can already see multiple threads forming—covering architecture choices, framework integrations, ethical concerns, security, webhooks, and more.

At Apexive, we've spent the last 12 months building AI-powered software with AI and integrating AI flows into Odoo modules (another good discussion on how to build Odoo modules with AI). We're happy to share our experiences with the community and collaborate to refine best practices.

The main goal of our repository https://github.com/apexive/odoo-llm is to centralize and structure AI development efforts around Odoo, preventing duplicated work and ensuring a productive, open-source collaboration. To keep discussions organized, I've also opened a Discussion section where we can create dedicated threads for specific topics without overwhelming this general conversation.

There are various approaches to hosting models and implementing agentic flows, and different Odoo users will likely have use cases for all of them. We've chosen Odoo as the foundation for this work because of its extensibility, modularity, and open-source nature. From what we’ve already seen with our customers in aviation and manufacturing, it’s a strong foundation for AI-driven business automation.

Looking forward to more insights and contributions from the community!

[ayushin]

I've create a list of topics above for convenience so we can go into separate threaded discussions:

• Security Considerations for LLM Agents in Odoo – @zamberjo
• Integrating Webhooks for AI Agents in Odoo – @rvalyi
• Standardizing AI Frameworks in Odoo (pydanticAI, LangChain, etc.) – @ValentinVinagre
• Hosting AI Models: Local vs. Cloud vs. Hybrid – @Tardo
• Defining the Role of AI in Odoo: Deterministic vs. Probabilistic Automation – @aleuffre
• AI-Powered ERP: A New Paradigm for Odoo? – @ayushin
• Ethical Considerations for AI in Open Source ERP – @aleuffre
• AI-Enhanced Odoo Automations: Use Cases and Best Practices – @rvalyi

[Charitablebusinessronin]

I'm considering https://github.com/coleam00/local-ai-packaged with https://github.com/tuanle96/mcp-odoo What’s included
✅ Self-hosted n8n - Low-code platform with over 400 integrations and advanced AI components

✅ Supabase - Open source database as a service - most widely used database for AI agents

✅ Ollama - Cross-platform LLM platform to install and run the latest local LLMs

✅ Open WebUI - ChatGPT-like interface to privately interact with your local models and N8N agents

✅ Flowise - No/low code AI agent builder that pairs very well with n8n

✅ Qdrant - Open-source, high performance vector store with an comprehensive API. Even though you can use Supabase for RAG, this was kept unlike Postgres since it's faster than Supabase so sometimes is the better option.

✅ SearXNG - Open-source, free internet metasearch engine which aggregates results from up to 229 search services. Users are neither tracked nor profiled, hence the fit with the local AI package.

✅ Caddy - Managed HTTPS/TLS for custom domains

[ayushin]

so local-ai-packaged seems like a bunch of tools pre-packaged together - not a fan, but this could work for you
mcp-odoo - is a just a connection between MCP and xmlrpc - you still need to build your AI super-app somewhere

odoo-llm is a toolbox for building AI apps and flows INSIDE of Odoo - rather then connecting external AI tools to Odoo

I believe that that's the right way of doing it - you need to bring the tools into your ERP, so the ERP becomes the center of all your automations and data (MCP host in the MCP terms)

We are still doing a few experiments but I expect the code to stabilise in a week time.

[ValentinVinagre]

It is true that the topic of "AI" brings up many questions. In essence, this module would allow Odoo to be connected to the "AI" that is required, so the biggest question here is: do we provide this tool to any user in the community? If so, according to OCA regulations, it is the user's own responsibility if they make improper use of this tool. Is this correct?
I think we should address this point first of all in order to direct our efforts to first bring a base module to the community and expand it.

[gbtechnology]

This is awesome! I hope we'll open a repository soon where we can start to contribute/work on this topic.

[ayushin]

@gbtechnology The repository is already there https://github.com/apexive/odoo-llm - we will transfer it to OCA as soon as it matures enough

[rvalyi]

seems legit, but then you better emphasize https://odoo-community.org/about/cla in your repo to avoid issues where third party contributions may be tainted later and it could not be possible to put it under the OCA umbrella. As AI is a hyped topic, better be cautious early with this...

[etobella]

I would prefer to move it ASAP in order to avoid CLA future issues

[ayushin]

Here is a quick explanation of what we have so far until our content team created something better:

https://www.loom.com/share/212def9456fd4103aba33ad0c1abe2d5

I've started another thread apexive/odoo-llm#80

What's your use case?

[arielbarreiros96]

Hi everyone!

I’d like to share a short video that demonstrates how a sales agent would work using the approach we’ve taken. In this case, the interaction happens through Livechat, but the same setup works seamlessly in Discuss as well.

sales_agent.mp4

Rather than integrating all AI tools directly into Odoo, we’ve opted for an external API architecture. The agent lives within this API, along with the tools it needs for each specific use case.

While we recognize the benefits of having a built-in automation framework inside Odoo, we believe those benefits don’t yet outweigh the downsides (especially when it comes to maintainability, support, and the ability to quickly introduce new capabilities). Given how fast the AI space is evolving, we think it would be extremely difficult to keep a newly created automation framework aligned with these rapid changes without sacrificing stability or development speed.

Additionally, it’s becoming increasingly clear that Odoo is developing its own vision of how AI fits into its ecosystem. For now, their focus seems to be primarily on task automation. While this approach has its strengths, we believe it complements our vision of integrating AI into Discuss chats, allowing us to enhance the platform’s capabilities and interactions.

That’s why we’ve chosen to decouple the AI logic from Odoo, which gives us the flexibility to evolve independently.

Looking forward to hearing your thoughts!

[arielbarreiros96]

Totally forgot to show the result lead, but the agent was capable of creating the lead with the resume of the request. This is also true for tickets and several other records the user may ask to create during a livechat session

[ayushin]

Here's some update on how odoo-llm can be used to build a chat bot using https://github.com/apexive/odoo-llm/tree/16.0/llm_knowledge

https://www.loom.com/share/6895612476e04a2e953442e5256aec97?sid=27be63c3-7aa4-4204-a81b-d686c4d7214f

[rvalyi]

@ayushin a dedicated AI repo has just been created in the OCA: https://github.com/OCA/ai
also read OCA/repo-maintainer-conf#94

[ayushin]

Not sure what's the idea here? @rvalyi Is this going to be a collaborative or alternative effort?

[etobella]

It is going to be a collaborative work. Anyone can publish and ask for review.

Now you can publish your code inside OCA if you want 😄

[ayushin]

OK, great. My only concern is time—I’m happy to contribute everything we’ve released so far in odoo-llm to the OCA (with attribution), but I can’t commit to the actual process. Would you be able to help with merging it?

Or perhaps it makes more sense to start a separate llm repo alongside ai?

Our approach is quite different—and, I think, more practical—than Odoo EE's ai* modules. We've focused on building agentic systems in Odoo, where assistants use encoded domain knowledge to perform complex tasks. It’s delivering strong results.

I do think these modules belong in OCA, but I’d hate to see the agentic direction diluted. Also, I feel Odoo's broad use of the “AI” label misses the mark.

[etobella]

@ayushin You can ask people to make the port, but I wouldn't recommend that. At the end, you did your own framework and style. It would be better to do it yourself. You can try to find someone to do the job for you... but I doubt that you will find this person. I would prefer that your team does the job, this way we can ensure that it keeps the original intent.

I would not recommend to create a second repository, at the end, we should simplify the number of modules on your repo. In the future, we can split them in separate modules if we have too many modules (in future versions).

Also, the name of the modules might be problematic, as you published them on apps.odoo.com. So, it will never be shown in OCA.

[bjoliveau]

100% agree @alexis A lot of confusion from Odoo between LLM, Agentic and automation. Bruno Joliveau Architecte solutions - Président +1 514-317-7944 <https://numigi.com/r/mwk> [image: LinkedIn] <https://numigi.com/r/AmB> [image: Twitter] <https://numigi.com/r/U6Y> [image: Facebook] <https://numigi.com/r/oNS> [image: YouTube] <https://numigi.com/r/xUI> Nouveau Tout savoir au sujet de l'Open Source Découvrir <https://numigi.com/blog/innovation-et-technologies-9/open-source-definition-avantages-66> Le dim. 8 juin 2025, 10 h 45, Alexis Yushin ***@***.***> a écrit :

…

OK, great. My only concern is time—I’m happy to contribute everything we’ve released so far in odoo-llm to the OCA (with attribution), but I can’t commit to the actual process. Would you be able to help with merging it? Or perhaps it makes more sense to start a separate llm repo alongside ai? Our approach is quite different—and, I think, more practical—than Odoo EE's ai* modules. We've focused on building agentic systems in Odoo, where assistants use encoded domain knowledge to perform complex tasks. It’s delivering strong results. I do think these modules belong in OCA, but I’d hate to see the agentic direction diluted. Also, I feel Odoo's broad use of the “AI” label misses the mark. — Reply to this email directly, view it on GitHub <#201 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJLN3VHJPIJ3WXMOD7R26T3CRD7RAVCNFSM6AAAAABX3S3CTCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNBQGI3TINY> . You are receiving this because you commented.Message ID: ***@***.***>

[ayushin]

well the subject of this post is HELP WANTED - read above - if anybody wants to help us port our modules into OCA contact me on linkedin please

[etobella]

@ayushin just a small comment. If you want to answer, use the Write a reply, do not open a new answer on the topic each time. Replying to a previous answer makes it easier to follow the story (future us will be happy with present us 😄 ).

Anyway, good luck finding people to do the port for you. You should remember that most people working in the community is paid by customers that asks for an specific functionality. At the end, we are on an OpenSource community. It is free, but not like free beer, but as freedom.

For example, if I find that a customer asks for something similar, I am not sure if I would use your repository. It contains a lot of extra things and it might take the same time (including understanding your structure, reviewing modules and porting) to port it to OCA than starting again with a simpler approach. Just once I found an external module that was introduced into OCA by other people (DMS module). In that case, the module had been there for 4 or 5 versions and a lot of people were using it. Also, it was ported on a Sprint Code, where people codes without an specific benefit. I hope yours is the second one.

[ayushin]

The reason the repository and module structure the way it is because we actually use it for different customers (and ourselves) in a number of different projects. It is not something built out of imagination, but things we use in production.

We've spent over year figuring out new and better ways of using AI - if you think you can do the same from scratch quicker and better, I'll be happy to use your code when it is ready.

[etobella]

I am sure about it, but an external developer doesn't know that. For this reason I said I would opt for something simpler. I was trying to explain the problem that developers will face.

You cannot expect that developers will understand your 29 modules easily and that they will follow your code by heart. The same will happen when other people review. It needs to be clean and easy to understand.

[ayushin]

thank you for your feedback @etobella - i'll discuss with my team to see what we can do about it.

[ayushin]

For those who want to run the reasoning loop outside of the Odoo - check out our new built-in MCP controller https://apps.odoo.com/apps/modules/18.0/llm_mcp_server

and
integration with MemGPT / Letta framework https://apps.odoo.com/apps/modules/18.0/llm_letta

[ayushin]

as soon as i can find some time!

[etobella]

Be aware that OCA is making their own way of handling it. The alignment will become harder when time passes.

[rvalyi]

The alignment will become harder when time passes.

prophetic!

[ayushin]

I can go ahead and open a PR to the OCA AI repository with the LLM modules as they are — would that help move things forward? @etobella

[etobella]

Well, you should need to follow OCA guidelines and enforce tests (I assume you already did, but I would like to confirm). Also, as there are other AI modules there, just check if there are some functionality clashes (same functionality). In that case, you should try to align your modules to work with the existent ones.

I know that it is some work, but this is the reason why we asked you to do it several months ago (before there were other modules)

[ayushin]

Here is the quick demo what you can do with the Letta & MCP Server:

https://drive.google.com/file/d/1VGlX1Q-EScaC3HKgtPTE8C5Dzbgw7BYy/view

[ayushin]

A better version of the above: https://youtu.be/fgHfyh2IZ04

[ayushin]

Finally getting some extreme value and ROI out of odoo-llm with https://apps.odoo.com/apps/modules/18.0/llm_tool_account

[clemenshelm]

This is the right gap to get right early, and the two risks above (data leaving to external APIs, and blocking destructive ops like model.unlink()) are really one problem: the model is non-deterministic, so you can't rely on it to decide to behave. The control has to sit outside the model.

What's worked for us: the agent starts with zero capabilities and each tool/action is allow-listed explicitly per agent. If unlink isn't enabled, the agent can't call it, no matter how it reasons. Add a tamper-evident audit log of everything it did call, keep irreversible actions behind a human confirmation step, and self-host with a local model for the data-residency concern. Then "is this safe" becomes a config + review question, not a trust-the-model question.

Disclosure so it's on the table: I build exactly this (Pinchy, AGPL, self-hosted on the OpenClaw runtime; it already runs governed agent actions in Odoo). I'm not pitching it as the answer here, the per-agent-allow-list + signed-audit pattern is reusable whatever framework you land on (PydanticAI/LangChain/etc). But since security architecture is explicitly on the help-wanted list, happy to help think it through if useful.

[ayushin]

I'm working on a permission first agent native execution environment that took a lot of inspiration from Odoo - https://angee.ai - it is about to go full open source next week.

[clemenshelm]

Nice, congrats on getting it to open source. Looking forward to seeing it next week, the permission-first framing definitely resonates.

Curious how you're mapping the permissions onto the agents themselves. Most agent tooling I've seen sticks to flat per-agent allow-lists, so going Zanzibar/ReBAC for agent authz is a more interesting bet. Would love to see how the relationship model plays out for non-deterministic callers once the code is open.

Either way, good to see more people taking the self-hosted, governed-by-default route seriously. Will dig in when it's out.

[ayushin]

@clemenshelm I sent you an email

[ayushin]

I'm looking for somebody ideally from OCA who has experience with both Odoo and Django https://angee.ai/careers/chief-ai-engineer-arpee/