SIMCONAPI Toolkit
API Toolkit
Working with agents · for engineers

You already know the moulding. This is how you put an agent to work on it.

If you have never worked with agentic tooling, you are not alone. Most of our customers had not either when they started with the toolkit. Here is everything you need to get going.

Nothing here assumes you write software. The valuable part you bring is the judgement about parts, materials and what good looks like · the work is getting that into a form an agent can act on, with you still deciding what counts as a good answer.

Write software already? There is a shorter version with the commands and the project layout.

The developer track
On this page
01What an agent actually isStart here if this is all new02Pick the program you run it inWhich to choose, what IT will ask03Your first sessionThirty minutes, with the prompts to paste04Keep the context written downYour agent offers to set this up for you05Think in verificationWhat the agent can check, and what you must06What usually goes wrong firstSymptoms, causes, and the fix07Learn from other people’s failuresCommunity sessions
01

What an agent actually is

An agent is a chatbot with one thing added: permission to act on your computer. It reads the files in a folder, writes new ones, and runs them.

So instead of telling you which simulations to run, it runs them and tells you what it found · much like a capable new colleague who does not yet know your parts.

Next

The program granting those permissions is called a harness. Choosing one is the next step.

Chatbot against agent
ChatbotAgent
Describes what you could doDoes it, then shows you
You copy code out and run itIt runs it and reads the result
Forgets between conversationsReads your notes every session
One answer at a timeHundreds of runs unattended
The loop you work in: you say what you need, the agent proposes a step, you approve it, it runs and reports, and the cycle repeats until the answer holds upYOUSay what you needin plain languageAGENTProposes a stepone at a timeYOUApprove itor stop it hereAGENTRuns and reportsreads the resultuntil the answer holds upTHE LOOP YOU WORK INYou never read the code. You judge the result.

You never have to look at the code it writes. What you do need is to read a result and say whether it is plausible.

02

Pick the program you run it in

A program that gives a language model hands · reading your files, writing new ones, running them · is called a harness. It is what turns a model into an agent.

We run Claude Code on a Max plan, so that is what we suggest. The field is wide and moving fast, and we would rather say what we have tested than rank tools we have not.

What we know as of 20 August 2026We update this as we test more.
Claude CodeWhat we runMax plan · VS Code or its own app

What we use ourselves, on a Max plan, and what we suggest you start with for the simple reason that it is the combination we know works with these recipes. It reads the CLAUDE.md and the skills that ship with every recipe, so it picks up the house rules and the platform facts without being told.

Get it
CodexWorked well for usVS Code or its own app

Same shape, different vendor · it reads AGENTS.md instead. We have had good results with it, though less recently than with Claude. A sound choice if your IT department already has an answer for it.

Get it
Cursor · Grok · Microsoft CopilotNot tested by usNo verdict either way

We have not put these through a Cadmould study in their current state, so we have nothing honest to tell you. Nothing in the toolkit depends on the harness · if you already work in one, tell us how it goes.

A sandboxed desktop app reads and plans fine but often cannot run the scripts, so the session stalls. Working from a project folder avoids that.

What IT will ask, and the honest answer · 3 questions
Where does our data go?The files the agent reads are sent to the model vendor to answer the request. Sent is not the same as trained on · business and enterprise plans exclude your content from training by default, and the vendor’s terms state the retention period. Check which plan you are on. Geometry and results stay local unless the agent uploads them to the platform, which it does over your own account.
Can it delete or change things?It can write files in the folder you start it in. Start in a fresh project folder, keep it out of network drives, and approve commands one at a time until you trust the pattern.
What does it need installed?Python 3.10 and outbound network access. The agent itself installs like any other developer tool. No admin rights beyond what Python already needs.
03

Your first session

Three steps to one finished result on screen. Start offline, so nothing can fail on the account side while you are still learning the tool.

Checklist0 / 3
Step 2 · paste this

Read the Get started page at toolkit.simcon.ai and set up the Cadmould API Toolkit in this folder the way it describes. Work through the steps in order, run the verify check after each one, and show me each command before you run it. Stop and ask me if a step needs a decision I have to make.

Step 3 · paste this

Open the bundled result. Tell me which region fills last and what the peak cavity pressure is. Show me the output you based that on.

The last line of that prompt matters. Asking for the evidence alongside the answer is the single habit that separates a useful session from a confident-sounding one.

04

Keep the context written down

An agent knows moulding in general and nothing about your shop · not the grades you run, not the machine the job goes on. Most of what it gets wrong in week one traces back to something nobody wrote down.

You do not set this up yourself · the skill offers to during setup. Say yes. It is an LLM wiki in the sense of Karpathy’s original note, and it compounds: a correction in the chat lasts until the chat ends, one in the notes holds every session after.

What your agent will ask

Set up context management for this project? I will create a notes folder for the grades you run, your machines and your house rules, and read it at the start of every session.

YesNot now

Say no and nothing breaks · the recipes still run. You just answer the same questions again next week.

05

Think in verification

This is the shift that takes longest, and it is worth naming early. Alongside the moulding question, there is a second question running: how would either of us know if this were wrong?

Tasks where the agent can check its own work go well and can be left running. Tasks where only you can judge need you in the loop. Sorting a job into those two piles before starting is most of the skill.

It can check these itself

  • Whether the code runs and the tests pass.
  • Whether a run completed and the result file opens.
  • Whether one candidate scores better than another on a stated metric.
  • Whether a number it wrote down matches the one in the file.

Give it work of this shape and it will grind through a thousand runs without drifting.

You still have to look

  • Whether the KPI measures what the customer actually cares about.
  • Whether the geometry it meshed is the revision you meant.
  • Whether a result is physically plausible for this part.
  • Whether the recommendation is one you would put your name on.

A confident summary is not evidence. Ask for the output behind it, every time.

Plan to experiment

Nobody has a proven recipe for this yet. Including us.

The technology is a few months old. Anyone selling you an established workmode is guessing, and the honest position is that the patterns are still being found.

So treat the first weeks as trials. Give the agent a task you could check by hand, see where it goes wrong, adjust the context, try again. Some attempts will fail outright, and those are the ones that tell you where the limits are.

You can skip a good number of those failures by hearing about them from someone who hit them last month.

See the community sessions
06

What usually goes wrong first

All of these have happened to us. None of them mean the approach is wrong.

SymptomWhat is actually happening, and the fix
It invents API calls that do not existIt is working from memory of similar libraries. Point it at the interactive reference and the skill that ships with the examples, and tell it to check the endpoint before writing the call.
It reports success without running anythingIt described the plan and stopped. Ask for the command output, or for the result file to be opened and a number read back. Make evidence part of the request.
It thresholds an absolute valueA reasonable instinct, and the wrong one for a fast screening sweep. Rank candidates against each other and let the numerical confirmation runs settle the absolute numbers. Put that in house-rules.md so it holds in every session.
It reads fine, then cannot run anythingA sandboxed harness whose permissions do not extend to executing the scripts. Nothing is wrong with your setup. Switch to Claude Code or Codex for now.
It keeps relearning the same thingThe correction stayed in the chat and ended with it. Fix it there, then tell the agent to write it into the wiki before it moves on.
It runs for an hour and produces a messThe task had no checkpoint it could verify against. Break it into steps that each end in something checkable, and look at the first one before letting it continue.
07

Learn from other people’s failures

We run regular community sessions where people working on the same problems compare notes: what worked, what wasted a week, which piece of context turned out to matter. Nobody has been doing this long. Everyone is a few months in, which makes the sessions unusually honest.

It is the fastest way in, and a good place to start even before you have installed anything.

Join the next sessionGo to the examples

Next: put a part through the solver

Setup is done. Hand your agent the first example and watch it run · one filling result, about nine seconds, on a part that comes with it.

First runAll examples