What changed lately, and what to watch for.
We release often, so this page moves often. What shipped sits next to what we know is still rough, because you should hear that from us rather than find it yourself.
AI filling simulations are queued now
A filling run used to come back in the reply to the request that started it, so a busy service could time that request out. Since Monday it is queued instead · submit the run, check its status until it is completed, then download the result. A run still comes back in a couple of seconds.
Anything written against the old reply has to be updated. On the Python SDK that is version 0.4.0 and one new wait step before the download; on the REST API it is submit, check, download; with the starter kit, replace your copy of the folder with version 0.2.0. The migration guide we emailed carries the exact code, and it is written to be handed to your agent.
What is new
newest firstDrive the toolkit from an MCP server
The MCP server is live at https://mcp.simcon.ai/mcp. Point an agent that speaks MCP at it and the toolkit becomes a set of tools it can call, so a workflow runs out of the conversation rather than out of a script you have to write first.
Calculate warpage on the numerical solver
Warpage runs on the numerical endpoint now, beside filling. So whether a part still holds its tolerances once it has cooled is a question you can put to the same endpoint you already send filling to.
Set up the toolkit from one prompt
A generator on PyPI writes you a working project, so the example folder you used to request from us is gone. Get started carries the prompt that does the whole setup, what has to be in place first, and what to do when a step fails · and it is the only place that carries it.
Filling runs are queued, so a busy service no longer times out
A run is queued rather than answered in the reply that started it · submit it, check its status, then download the result. You can cancel one while it waits, and a run still comes back in a couple of seconds.
Every AI run now goes through the new Humboldt model
Humboldt is the model behind every AI run. Against the model it replaces it is a significant step in both accuracy and performance, and there is nothing to turn on · the runs you already send go through it.
Ask whether a part can be moulded, from the geometry alone
Nine checks now run on a geometry on its own · wall thickness, draft angle, flow path, air traps and the rest · so a part a customer sent this morning gets an answer without a simulation being set up first. Every part we have measured came back in under ten seconds.
Run filling on the numerical solver, not only the AI one
Filling can now be run through the numerical solver as well as the AI one. The AI solver stays the fast way to compare many options against each other · the numerical solver is the one to reach for when a result has to be confirmed rather than ranked.
Open any result in the web viewer, straight from the run
Every run now comes back with a link that opens that result in Cadmould Web. If looking at it is all you need, there is nothing to install and nothing to write.
Set a simulation up in the browser, not only in code
The platform has a graphical way to configure a run · choose the geometry, the material and the process settings and start it, without writing the call yourself.
Known limitations
1 currentLarge parts sometimes go missing in the web viewer
On a big part the camera can end up inside the geometry, and anything nearer to you than it can focus on is not drawn. The part looks cut open, or disappears as you move around it. It does not show on every part or from every angle, so one clean run is not proof you will not meet it.
Zoom out until it comes back. Nothing is wrong with the run · the result file holds the full part, and every number you read off it is unaffected.
Hit something that is not listed?
Tell us what you were running and what you saw. A fault we know about goes on this page, so the next person reads it here instead of losing any time to it.