Please find below the questions that I had no time to answer during the webinar, and the corresponding answers!
What is the pricing model for the AI for Capella plugin?
AI for Capella is a commercial add-on, licensed as a subscription per named user (floating options coming). Please contact Obeo for a quote. A free one-month trial is available. Note that Capella itself remains free and open source, and you use your own LLM.
When you give an LLM access to a Capella model, what do you expose: the raw XMI, a Python API like capellambse or py-capellambse-based tools, or a curated semantic view? What did you learn about which works best?
Neither raw XMI nor a generic API alone. The MCP server exposes a curated semantic layer: tools that speak Arcadia/Capella (explore architectures and interfaces, trace functional dependencies, analyze impact, run Arcadia transitions, create elements the way Capella’s own wizards do). Generic model access through queries is a fallback. Our experience: semantic, task-level tools give much more reliable results than letting the LLM navigate the raw model structure, and they consume far less context.
Diagram format or layout is a significant task for modelers today. Would AI for Capella allow AI to perform rearrangement of diagrams?
Yes. The agent can already create and arrange diagrams by calling layout algorithm available in Capella, including ELK layout if installed (available with Capella 7.1). We are working on using AI to layout diagrams as well, with very promising results, but this is not yet available.
Isn’t this more of a demonstration of the LLM’s power than of the Capella tool and the ARCADIA methodology?
Both, in fact. The LLM brings language understanding, but the quality of the result comes from Capella and Arcadia. The method gives the agent a structured space to reason in. Capella’s semantics (capabilities, functions, exchanges, functional chains) give it meaningful concepts. And Capella’s own operations and validation keep the model consistent. The same LLM without a structured model and method produces far less reliable engineering results.
To what degree is AI for Capella Open Source so it can be used in a local systems engineering ecosystem that requires maximum privacy, using locally installed and trained LLMs.
AI for Capella is a commercial add-on, not open source, but it is designed for private environments. It is installed on-premises in your Capella, connects through standard MCP (stdio, HTTP, HTTPS), and works with the LLM of your choice, including locally hosted models. No data has to leave your network. Read-only and edit modes let you control what the agent may do.
The paradox is that the ARCADIA methodology is designed for humans who struggle to manage complexity, not for an LLM. So is the next step the disappearance of the methodology itself?
We don’t think so. Arcadia doesn’t just help humans manage complexity. It structures engineering reasoning, decisions and responsibilities, and that is exactly what makes AI output reviewable and trustworthy. Engineers still decide and validate, and they need a shared, structured model to do so. The method becomes even more important as a guardrail for AI.
Est-ce que on peut configurer les diagrammes que on souhaite créer ?
Oui. Vous indiquez à l’agent le type de diagramme souhaité (par ex. LAB, SDFB, OES…), le périmètre et les éléments à afficher. Il crée le diagramme via les outils Capella, et vous pouvez ensuite l’ajuster.
What functionality is available during the demo phase of the plugin, and how long does the evaluation period last?
The trial gives you full features for one named user, deployed on-premises (non-production use), for one month, with access to Obeo’s support team. Contact Obeo to request it.
Can help to generate a document from the model with same Capella views and traceability Matrix please?
Yes. The agent can generate and enrich documentation from the model and export diagrams, with traceability preserved. For formal, repeatable documents (Capella views, traceability matrices, company templates), we recommend combining it with M2Doc, Obeo’s open-source Word generator for Capella models.
How does your approach ensure that the MCP server enables genuine reasoning over Capella’s engineering semantics, independently of the chosen LLM, rather than merely providing an uncontrolled automation layer that navigates the model’s structure?
The MCP server doesn’t expose the model as an unstructured graph to crawl. It exposes Arcadia-aware operations (transitions, impact analysis, functional dependency tracing, element creation through Capella’s own rules) and returns structured engineering context. Capella itself applies and controls every change, so model consistency doesn’t depend on the LLM. Plans are proposed to and validated by the engineer. Better LLMs reason better, but the guardrails are the same whatever model you use.
Can’t we integrate Copilot with Capella?
Yes. Any MCP-compatible client can connect to AI for Capella. This includes GitHub Copilot (agent mode in VS Code) and Microsoft Copilot Studio, both of which support MCP.
Have you tested the replicability of the results? Did you manage to reach a coherent/similar outcome when giving the same input prompt?
LLM outputs are not fully deterministic, so two runs of the same prompt can differ in wording or detail. In practice, results are consistent in structure and substance when the request is precise. Variation is limited by the deterministic MCP tools, which give the same answer for the same model, by Capella’s own transformations, and by the plan-validate loop. For repeatable tasks, we recommend reusable prompts or scripted workflows.
How much information of the system of interest (or related systems) does it need to obtain results with sufficient level of quality?
It depends on the task. Querying, explaining and impact analysis need only the model itself. For model creation, the quality of the input matters most: clear needs, a mission and capabilities, existing product documentation or requirements. Starting from structured sources and working step by step (one Arcadia perspective at a time) gives much better results than a one-line prompt.
Is this available for SysON
Yes. AI capabilities for SysON are available in Obeo Enterprise for SysON.
In the past, I have been blocked by the pricing of Obeo products versus alternatives. The closest comparison I found is: Celedon Davinci (Core: $20/month/user, Max: $200/month/user), free viewers. Dalus.io (Basic: $99/editor/month, Pro: $249/editor/month) with near unlimited free non-editors with Copilot access. How does this compare to the pricing of Capella + Publication + AI (which gives a similar result)
Thanks for the comparison. We don’t publish prices, so please contact Obeo for a quote, but we are competitive when considering these prices. Keep in mind the overall picture: Capella and Arcadia are free and open source, with no lock-in, and AI for Capella lets you use your own LLM and AI environment rather than a bundled one. A free one-month trial lets you assess the value on your own models.
If modelling in Capella / Arcadia using AI is that easy. Are you not afraid that all the business around Capella / Arcadia training will fall ?
Not at all. AI makes it faster to produce a model, but it doesn’t remove the need to understand the method, judge architectural choices and validate what the AI proposes. Engineers need to master Arcadia to direct and review the agent. If anything, training evolves: less tool clicking, more method, modeling judgment and AI-assisted workflows.
Sorry I didn’t understand whether I can go back through the agent’s changes one modification at a time on my model. Can we roll back the changes?
Yes. Each operation performed by the agent is executed as a Capella command, so you can undo the changes step by step with the standard Undo, just like manual edits. For larger rollbacks, use your usual versioning (Git, Team for Capella). The plan-validation step also lets you review changes before they are applied.
Est-ce que svp, on peut seulement faire une génération de la couche opérationnelle seulement ?
Oui. Vous pouvez demander à l’agent de ne travailler que sur l’analyse opérationnelle (entités, acteurs, capacités, activités opérationnelles…), puis d’enchaîner sur les autres niveaux seulement si vous le souhaitez.
How do I connect AI with Capella? Do I need to install any add-ons or configure specific parameters? Also, how can I make AI automatically build models in Capella? The demonstration just now went a bit too fast, and I didn’t catch everything—could you please explain it again briefly
In short: (1) install the AI for Capella add-on in Capella (7.1, 7.0.1, 6.1 or System Modeling Workbench; Java 21), (2) configure the MCP server (stdio, HTTP or HTTPS, read-only or edit mode), and (3) connect your MCP-compatible AI client (Claude, ChatGPT, Codex, Gemini, Copilot…). Then just describe what you want in natural language: the agent proposes a plan, you validate it, and Capella applies the changes. The webinar recording is available on the Capella YouTube channel, and the 2026.11 release will bring a simpler installation and short configuration videos. Contact Obeo for a trial.
Which input did you provide to the agent to teach and apply the ARCADIA methodology? Particularly, did you provide more structured methodology descriptions beyond the user handbook, tutorials, text books and examples?
Current LLMs already have good knowledge of Arcadia from public sources. On top of that, the MCP server embeds structured guidance (Arcadia steps, the right Capella operations and transitions to use), which we will keep improving in future releases.
Does it work with Ollama for using private local LLMs? Thank you
Yes. AI for Capella works with any MCP-compatible client, including clients using local models served by Ollama. See the next answer for our experience with local models.
What is your experience of running AI for Capella with local LLMs? Context size? Speed? Quality?
It works, but quality depends on the model. You need a model with reliable tool calling and a large context window, because model exploration can return a lot of context. Smaller local models tend to be slower and less reliable on multistep tasks. Larger open-weight models give good results on querying and analysis, and model creation is more demanding.
Great to see this progress. Can only hope it doesn’t make us all lazy.
That’s a fair concern! That’s why we designed AI for Capella around human control: the engineer requests, validates the plan, and validates the result. AI should take away the tedious work, not the thinking.
Does it work with local LLM engines such as Ollama? Thank you
Yes. See the answer above: it works with any MCP-compatible client, including local LLMs served by Ollama.