ParaCrux

How the Agent decides

Every answer has a source. Every change has a gate.

What the Agent reads, what leaves your workstation, how each request gets routed and where you stay in charge. Plain facts, including the parts that aren't finished.

YOUR WORKSTATIONlocal context engineHOSTEDglobal context engineSHARING ON TO CROSSmessage + packetshadowprecedentsone proposed callruns on a copy · original keptSOLIDWORKS sessiontyped inspectionDesign contextper-workspace storeDesign libraryone design · workspaceReference filesname + fingerprint onlyContext builderallow-list · no pathsYour reviewworksheet → copyReasoning model1 tool call per stepDecision layertyped · abstains < 80%Engineering memoryprivate + paid commonsIN STAGING

Local context engine: your workstation

Live

The Agent starts from what is in front of you. It reads the open SOLIDWORKS model through typed inspection: features, dimensions, configurations and properties.

It adds a per-workspace design-context store and your design library, scoped to one design or shared across your workspace. Search here is literal and bounded, not guesswork.

Reference files never leave. The model sees their names and fingerprints only. A context builder assembles a small, allow-listed packet and strips file paths.

What leaves your machine.

No design context leaves until you turn on design-context sharing in the desktop app. Then only the right-hand column does.

Stays on your workstation

  • Your model files and assemblies
  • Reference file contents: PDFs, images, CAD
  • File paths, which are stripped from tool results
  • Your running SOLIDWORKS session and licence

Sent when sharing is on

  • Your message
  • An allow-listed packet of design context
  • Results of supported tool calls
  • Reference files as names and fingerprints only

Managed AI uses OpenAI with global processing. See the privacy notice.

From request to result.

Six steps, in this order, every time. The two that matter most are yours.

  1. Read the request

    Your message, plus a small allow-listed packet of design context built on your workstation. File paths are stripped first.

  2. Route it

    In shadow mode today: the decision layer records which typed route it would choose, and how confident it is. Below 80% it abstains. It doesn't steer the Agent yet.

  3. Gather context

    Local first: the open model, your design context and library. Hosted engineering memory, now in staging, only when it's enabled for you.

  4. Propose one step

    At most one tool call per step, up to eight steps and £2 per job. Credit is reserved before each step and unused credit is returned.

  5. You review

    Dimension changes land in a worksheet. Nothing runs until you approve it.

  6. Run on a copy, then check

    The original is kept. Native and STEP outputs are there to inspect, and each managed step gets a receipt.

A model that doesn't write.
It chooses.

Language models are good at prose and bad at admitting doubt. So routing goes to a separate typed model: fixed options, a confidence score, and permission to say “not sure”.

REQUEST

“Read and inspect my existing SLDPRT reference. Do not change it.”

A real request from our own use; file name removed.

ROUTES IT CAN CHOOSE

  • inspect
  • create box
  • profile search
  • clarify
  • unsupported

RESULT

Decided: inspect

Clear request, clear route. The Agent would inspect and change nothing.

From our routing benchmark of 28 September 2026: 16 cases, each run twice. 26 of 32 answers cleared the 80% floor and all 26 matched the expected route; 6 abstained. Median answer time 270 ms. A small, deliberate test set, not a production accuracy figure, and the layer runs in shadow mode.

What it won't do.

  • Save over your original. Changes run on copies.
  • Run a change you haven't reviewed. Dimension changes wait in the worksheet.
  • Upload your model files. Only what the right-hand column above lists.
  • Retry a CAD step on its own. A failed step is never silently replayed.
  • Keep going when an outcome is uncertain. The loop stops instead.
  • Spend past its limits. Each step reserves its cost first; a job stops at £2.

Credits, plainly.

Each paid seat includes £5 of managed AI credit every UTC month. It is spent first and refills on the 1st. Top-ups of £10, £25 or £50 are spent next and carry forward.

  • Your dashboard shows both balances as meters, with the refill date
  • Each step reserves its maximum cost, then returns what it didn't use
  • A job stops at £2 or eight steps, whichever comes first
  • Rates are listed in the desktop Agent settings and the terms
ILLUSTRATION · ONE SEAT, ONE £25 TOP-UPNOT YOUR ACCOUNT
  1. Included credit is spent first.
  2. Then your top-up credit.
  3. On the 1st, included credit refills. Top-ups carry forward.

Is it unreasonable to try it on a copy of a part you already have?

The Agent works on copies, so a trial run leaves your original untouched. Check compatibility first; we'll tell you straight if your version or workflow isn't a fit.

Find your plan See the lab recording