Dolibarr Skill (AI Agent)

Developing for Dolibarr
, , ,
1

Hi,
if you’re interested I shared the very first Dolibar developer skill for your AI Agent
npx skills add jeffsenso/dolibarr-skills

For usage of skill, check the web or skills.sh QA

It’s based on Dolibarr official documentation.
Your agent don’t guess anymore, it knows :wink:

1 Like
2

Hi Jeffson, is this a new module or a built-in configuration?

3

SKILLS are the new superpower of the Agents. :flexed_biceps:t4::flexed_biceps:t4::flexed_biceps:t4:

Briefly: a plain text file (or a collection of them, even including templates or scripts) installed in a hidden directory of your project you are working on, helped by an AI agent (it can also be “installed” in the root settings directory of the agent itself).

Note: when i say “installed” it means a simple COPY the files! Amazing!!

These SKILL files provide POWERFUL, STRUCTURED, and ACTION-FOCUSED rules and documentation (even code tools) to make the AGENT become a “super-expert” in a task or environment. In this case: “developing inside Dolibarr according to Dolibarr standards & rules.”

Just now (first quarter of 2026), this new “agent AI feature” is exploding and has been changing 100% the ability and quality of the agents’ performance in such an EASY WAY that all of us developers are still not able to believe it. This will change a lot of things in 2026.

Wonderful forward jump!! :rocket::rocket::rocket:

4

Just in case this is useful for any other user.

Tech setup:

  • I’ve installed opencode as the AGENT engine (open source).
  • I’ve installed it in a local VM (Ubuntu on Ubuntu, using multipass), with 4 GB RAM and 2 CPUs.
  • The trick is to HAVE a shared directory on the HOST machine (mounted inside the VM). This magic is done by the virtualization engine. In my case, it’s very simple:
    multipass mount ~/Documentos/_OPENCODE_ opencode-vm:/home/ubuntu/proyectos
    In this way, the agent software (opencode in my case) can work on all the projects I put there, and I can also test and take the output files of the agent’s work from the HOST’s original directory. Perfect! And the agent cannot look outside that directory. It’s a barrier shielded by the virtualization system itself :blush:.
  • This setup lets me have the AI Agent working in a “controlled jail,” so I can install any other tool or script there: to process files, a CLI web browser (to give the Agent eyes to navigate websites, even sessions… to test your module inside your local Dolibarr :grin:), open-source OCRs, open-source TTS, etc.
  • Inside the VM i use tmux new -s opencode to 1) maintain an open and run session with opencode, and to 2) be able to open new “shell panels” were to install things, change permissions, etc.
  • In fact, i start opencode web to have a web UI. For that i added this alias to the file ~/.bash_aliases and put this:
    …
    alias ocweb='OPENCODE_SERVER_USERNAME="myownusername" OPENCODE_SERVER_PASSWORD="myownpassword" opencode web --hostname 0.0.0.0 --port 4096' … and voilà!! :smiling_face_with_sunglasses:

Regarding daily use:

  • Honestly, I’ve been using it for 15 days now, and it’s really amazing :star_struck:.
  • Once you start interacting with the agent, you can’t stop opening new projects every day.
  • With every new project you do using the agent, you usually discover new SKILLS (you should) to be “installed” (used).

Some of the amazing uses I’ve had so far:

  • Create an n8n workflow from scratch (the agent guided by me). Tech note: by installing MCP or API credentials, you can give the agent access to remote services (like your n8n VM, etc.). n8n “recently” has an MCP server by design, which you can call from an agent by simply enabling this setting! Then n8n lets the agent create/edit/run/test/monitor the workflow and its executions. Recommended: install a SKILL to work with n8n beforehand :smiling_face_with_sunglasses:. It’s optional, but your agent will do a much better job.
  • Web search for a certain topic, extract a report, save it to local Markdown files, and then generate: beautiful HTML dashboards, PDFs, interactive tiny apps, etc.
  • Given a webpage, the agent can “see it” and generate a second version with improved SALES COPY. Recommended to use a SKILL for this.
    :star: I love this: landing-page-copy — rampstackco/claude-skills
1 Like
5

Great thread — this addresses something I think a lot of developers hit: Claude or Cursor “knowing” Dolibarr’s architecture is very different from understanding your specific instance.

@Jeffsenso’s skills package handles the framework layer — conventions, hook patterns, module structure. @caos30’s VM sandbox solves the safety layer — giving the agent a controlled environment to act in without risk to production data. Both genuinely useful.

One layer worth adding to the stack: retrieval. Skills + schema knowledge answers “how do I build X in Dolibarr,” but a RAG layer over your actual data answers “what does my instance say” — so the agent can reason over your invoices, client history, and stock movements, not just the framework. I’ve been building this combination for production Dolibarr deployments; the two layers handle quite different question types and complement rather than replace each other.

For anyone testing on a live instance: even outside a VM, pointing the agent at a read-only API key first limits blast radius until you’ve mapped what write operations the agent can reach.

— Ali — Dolibarr AI Consultant / SiliconBlaze.com