cadloop

A closed loop from parametric source to G-code, with the verification step in the middle that neither the CAD tool nor the slicer can do.

Two MCP servers. A model writes OpenSCAD, compiles it, reads computed values back out, looks at a render, checks it fits the bed, slices it and pulls out the G-code, with every plate proven against the printable area. No GUI at any point.

The verify step in the middle is yours to write. cadloop makes everything around it cheap enough that you get to run it often; it cannot tell you whether the part works.

Every number in that walkthrough is real output. The failing wheel is a real measurement, not an illustration. How it's built and where each number comes from is in docs/walkthrough.

 setup_printer   once, no arguments: which printer is this for
      |
  write .scad
      |
   check          syntax and references, about a second
      |
   echo           read computed values without rendering
      |
   preview        look at it
      |
   verify         does it actually work  <- yours; nothing here does it for you
      |
   measure        bounding box and volume
      |
 check_bed_fit    against the printer you set up
      |
  slice_model     -> .gcode.3mf
      |
 extract_gcode    -> .gcode

The ordering is the point. check and echo skip geometry evaluation entirely, so they cost a second against models where a full render takes minutes. Pay for the expensive steps only once the cheap ones pass.

Everything from the bed check down is also one call, make_printable, which renders, packs, slices and proves the result on the bed and reports what it did in a screen of text. See the worked example.

Install

pip install cadloop

The only dependency is mcp. Or from a checkout, which is what you want if you intend to run the worked example or the tests — the verify extra is shapely, needed by the example's own checker rather than by cadloop:

git clone https://github.com/richardofortune/cadloop
cd cadloop
pip install -e ".[verify]"

Needs OpenSCAD on PATH or at OPENSCAD_BIN, and for slicing, OrcaSlicer, Bambu Studio, ElegooSlicer or Creality Print. They share a CLI, so any of them works, and all four are auto-detected along with the profiles they ship.

No slicer is preferred over another here. setup_printer probes each one it finds and then proves the winner by slicing a 20mm cube with your profiles, so which one you get is decided by which one works on your machine rather than by the order of a list in this repository. Creality Print's CLI is broken headless on macOS today, and that is something this discovers rather than something it asserts, so it corrects itself when upstream ships a fix; see testing status below. SLICER_BIN overrides the choice.

Configure

Copy mcp.json into your client's config and fix the paths. Both servers should point at the same workspace directory, which is the only place either one reads or writes.

The servers

openscad exposes check, echo, render, measure, preview and workspace file access. preview returns the PNG as an MCP image, so the model can look at what it built rather than inferring from numbers. render returns OpenSCAD's manifold report alongside a bounding box and volume, where simple: yes with a sensible volume count is the signal the mesh is printable.

The two split on what happens to the picture. preview hands it back and keeps nothing; render writes it to the workspace when the output ends in .png, taking the same camera, imgsize, projection and full_render options. Use render when the image is the artefact rather than the answer. Since render also takes source as text rather than a path, the before half of a comparison is git show <rev>:model.scad piped straight in — no checkout, no temp file.

slicer exposes setup_printer, machine_info, make_printable, slicer_info, list_profiles, check_bed_fit, slice_model, slice_summary and extract_gcode.

make_printable is the whole chain in one call. Give it a .scad and the parts you want out of it, and it renders each one, measures it, packs what fits onto as few plates as it can, slices them, and then reads the finished G-code back to prove every extruding move lands on the bed. It never edits the model: a part that cannot print as designed is reported, not quietly shrunk or split. The one thing it changes is which plate a part lands on, and turning a part ninety degrees when it will not fit square, which it names in the report. It answers with every fact it established and, in summary, the one screen of text those facts add up to, ending in what to do next.

Call setup_printer() once, with no arguments. It reads what your slicer is already configured with, resolves that to a machine, process and filament profile from one install, proves the combination by slicing a 20mm cube, and remembers it. It reports the printer, the quality and the filament it settled on, and stores nothing at all if the test slice fails. Pass printer, filament or process only to override one of those fields; the rest still come from your slicer's own settings. A profile you name is either used or refused by name, never quietly swapped for a different one.

After that, check_bed_fit and slice_model need no profile arguments, which is the point: a caller cannot supply three profiles that disagree if it supplies none. Explicit arguments still win where you pass them.

machine_info says which printer this workspace is set up for and whether it is still current. The record is a cache, not a source of truth. If the slicer moves, a profile is edited, or the slicer's version changes, every tool that would have used it refuses with the reason instead - ok: null, nothing written - until you run setup_printer again. A setup that no longer matches reality is never used to produce G-code.

Call slicer_info before doing anything unusual: the Orca-family CLI is undocumented, changes between releases, and Creality's fork diverges, so the flag list it reads off your install is more trustworthy than anything assumed. slice_model has an extra_args escape hatch and a dry_run mode.

check_bed_fit is where the two meet. The slicer will emit out-of-bounds G-code without complaining, so this measures the STL, reads printable_area out of the machine profile, and compares. It checks the 45 degree diagonal too, since a part that misses square-on often fits rotated.

Reading printable_area back out is fiddlier than it looks. Stock profiles write it both as a list of "XxY" strings and as one comma-separated string, and most Bambu machines carry no bed of their own at all, inheriting it through inherits from a common base. So the lookup parses both forms and walks the inheritance chain. Of the 473 concrete machine profiles shipped with Creality Print 7.1.1, 469 resolve; the remaining four define no bed anywhere in their chain, and those report ok: null with a reason rather than guessing.

The verification step

This is the part cadloop does not do for you, and the reason it is a worked example rather than a feature. models/verify_spirograph.py ships beside the model, not inside the package: every project's version of "does it actually work" is different, and none of them generalise. Yours will look nothing like this one.

It runs two checks. For the spirograph the first lays each wheel's pitch curve onto the ring's pitch circle, walks a full circuit, and measures overlap between the two solids at every position. Zero overlap across the whole circuit at some meshing phase is the pass condition. The second checks the parts are not laid on top of each other on the sheet.

$ python models/verify_spirograph.py
part     teeth  overlap mm2  result
24T         24     0.000000  pass
...
trefoil     23     0.000000  pass

14/14 parts mesh cleanly

group        volume mm3
ring            23939.8
outer_ring      25652.8
wheels         197113.5
shapes          10078.7
sum            256784.8
sheet          256784.8

no parts overlap on the sheet

The second check is the same idea one level up. A sheet that lays two parts on top of each other still renders as a clean manifold, still fits the bed, and still slices without a word; it just prints as one fused object. So it renders the sheet and each of its groups and compares the union against the sum, which is the only place the collision shows up. It needs OpenSCAD and skips without one; --skip-layout and --skip-mesh run one half alone.

What the tooth counts decide

A pen in a wheel of r teeth rolling in a ring of R traces a figure with R / gcd(R, r) lobes, closing after r / gcd(R, r) circuits. The pattern is settled by the tooth counts before any geometry exists, so the checker's --patterns mode reads it straight off the set:

$ python models/verify_spirograph.py --patterns