KOMPAS-3D gets an AI agent that builds parts, drawings, and exports DXF on its own

KOMPAS-3D has received an AI agent that builds parts, drawings, and exports DXF on its own

The MCP for KOMPAS-3D took a path over two months from an impressive demo to an agent that can be trusted with real CAD routine. Now AI doesn't just send commands to the system, but builds a part itself, tracks the model state, saves the document, and returns the result with a screenshot.

From Demo to Practice

The first version of the bundle looked interesting, but was rather a proof of concept. The skepticism was understandable: if the agent doesn't understand what stage the model is at, any error in the sequence of actions breaks the entire chain. For an engineering environment, this is enough for the tool to remain a toy for presentations.

That's why the main update wasn't another command, but the ability of AI to maintain context of work within a specific document and work session. Now the scenario looks noticeably more mature. The agent receives a task, creates a 3D part, adds holes, checks the build tree, saves the file and generates a screenshot of the result without manual intervention between steps.

In parallel, it can take on part of the flat routine: generate a drawing, fill in the stamp and export the result to DXF in a single request. This is already closer not to "wow-demo", but to a tool that can be tested in a real workflow.

"Cool for a demo, but won't fly in real work" — this is how the author

describes the reaction to the first version of the MCP.

Memory of the Model

The key difference in the new version is that the agent keeps in memory the current state of the model at each step. It understands which document it's working in, what base body has already been created, how many elements are hanging in the tree, and which feature appeared after the last operation. Because of this, AI doesn't act blindly and doesn't try to rebuild the part from scratch every time if work has already been started before.

The article provides an example with get_3d_context: the tool allows you to ask the already open part about its current state and immediately see that the base geometry is ready, and in the tree, for example, there are already 11 elements. This eliminates guessing and makes the next operation a logical continuation of the previous steps. For CAD tasks this is critical, because the cost of a wrong assumption here is higher than in an ordinary text agent or office scenario.

The second important layer is working with object selection in the scene. Instead of manually calculating coordinates and snap points, the agent can descriptively find the needed face through resolve_selection_3d, get its system identifier and use it for the next operation. If the top flat face is needed, AI finds exactly that one, sets a new sketch and makes a cut there where it is actually required by the model, not by approximate heuristics.

Verification and Documents

Equally important is the verifiability of the result. Commands no longer "fly into the void": the list_feature_tree_3d tool returns the build history with specific steps like base sketch, extrusion, cut, hole or chamfer. This means that an engineer can open the document and verify not only the final shape of the part, but also how exactly the agent arrived at it. For production software, such transparency is often more important than flashy automation and beautiful demonstrations.

• Building a 3D part from a text description
• Adding holes and other features in the correct sequence
• Creating a drawing, auto-filling the stamp and exporting to DXF
• Saving the document and returning a screenshot of the finished result

Essentially, the MCP for KOMPAS-3D begins to cover two levels of work at once: volumetric geometry and accompanying documentation. This is especially interesting for typical operations where a specialist spends time not on engineering solutions, but on repetitive actions in the interface. If the agent stably maintains context and knows how to confirm each step through the build tree, it has real chances to take a place alongside familiar CAD automations.

What This Means

The story with KOMPAS-3D shows that the usefulness of an AI agent in engineering software is determined not by beautiful dialogue, but by the ability to remember the state of the model, correctly select objects and leave a verifiable trace of actions. If such integrations become more reliable, the next wave of AI in CAD will not come through chatbots, but through practical tools for parts, drawings, export operations and typical engineering routine directly within familiar CADs and team processes.