Cadloop MCP Server

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 and a checker. A model writes OpenSCAD, compiles it, reads computed values back out, looks at a render, proves the parts actually work, checks they fit the bed, slices them and pulls out the G-code. No GUI at any point.

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.

  write .scad
      |
   check          syntax and references, about a second
      |
   echo           read computed values without rendering
      |
   preview        look at it
      |
   verify         does it actually work  <- the part tooling usually skips
      |
   measure        bounding box and volume
      |
 check_bed_fit    against the machine profile's printable area
      |
  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.

Install

pip install "cadloop[verify]"

Or from a checkout, which is what you want if you intend to run the worked example or the tests:

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. Orca and its forks are preferred over Creality Print, whose CLI is broken headless on macOS; 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.

slicer exposes slicer_info, list_profiles, check_bed_fit, slice_model, slice_summary and extract_gcode. Call slicer_info first: 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

cadloop-verify is the reference example of the third thing you need, and the one no general tool provides. 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.

$ cadloop-verify
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 cadloop-verify --patterns reads it straight off the set:

$ cadloop-verify --patterns

main ring, 96 teeth
      part  teeth  lobes  circuits
       32T     32      3         1  plain
       24T     24      4         1  plain
       72T     72      4         3  plain
...
   trefoil     23     96        23

96 is 2^5 x 3, so it shares factors with most of an even wheel set and a good half of these wheels draw eight lobes or fewer. There is no 48 in the set for this reason: 48 is exactly half of 96, the degenerate ratio whose pen traces an ellipse and nothing else, so it earned no slot. The outer ring, 105 = 3 x 5 x 7, behaves far better, and the trefoil at 23 teeth is coprime to both rings, which is what makes it the richest wheel in the set.

This is worth running before choosing tooth counts rather than after printing them.

This is model-specific by nature, which is the honest lesson. A manifold mesh that fits the bed and slices cleanly can still be a part that does not work. Whatever your equivalent of "does it actually roll" is, it belongs in the loop between preview and measure, and you have to write it yourself.

The worked example

models/spirograph.scad is the model the loop was built around. A 96 tooth internal ring in a flanged body, an outer ring, eleven circular wheels and three non-circular ones (ellipse, egg, trefoil), all involute geared at module 1.5 with pen holes on a golden-angle spiral.

The non-circular wheels are why the verifier exists. A first attempt used capsule and teardrop outlines built from tangent lines; those looked right, rendered as clean manifolds, and would have sliced without complaint, but the rolling check showed them ploughing 30 to 80 mm² into the ring teeth. A flat section of pitch curve touches the ring at one point and stands proud of it everywhere else. Nothing downstream of CAD would have caught that. The shapes that shipped are smooth convex curves whose radius of curvature stays inside the ring's everywhere.

make verify     # rolling interference and layout, all 14 parts
make render PART=ring
make smoke      # both servers, end to end

Testing status

make smoke drives both servers over real MCP stdio sessions and asserts on twenty-one behaviours: tool surface, defines reaching the script, a deliberate syntax error being caught, measured geometry matching known dimensions, an image coming back from preview, profile classification, argument ordering, archive parsing, G-code extraction, bed fit passing and failing, both printable_area spellings and an inherited bed, and the workspace guard rejecting traversal.

The OpenSCAD half runs against a real OpenSCAD and skips if none is installed. The slicer half runs against a mock binary that emits an Orca-shaped help text and a representative .gcode.3mf, so everything except the real slicer's own behaviour is covered.

Against a real install, the OpenSCAD half and check_bed_fit are confirmed end to end: a 24 tooth wheel rendered out of spirograph.scad measures 38.99 mm and passes the K1's 220 mm bed. Every flag slice_model builds is present in Creality Print 7.1.1's --help.

The loop is confirmed end to end against OrcaSlicer 2.4.2 on macOS: the 96 tooth ring slices to 113,808 lines of G-code over 35 layers, a 51 minute print using 7.27 m of PLA, with every coordinate inside the bed.

Creality Print's CLI does not work headless on macOS. On 7.1.1.4472 under macOS 26, every operation except --help segfaults, including --info with no profiles loaded at all. It dies in Slic3r::GUI::PartPlate::set_shape called from Slic3r::CLI::run, a null dereference in GUI bed setup that the headless path never initialises, and no combination of --datadir, --outputdir or profile arguments avoids it. Point SLICER_BIN at OrcaSlicer, Bambu Studio or ElegooSlicer instead, or run Creality Print's CLI on Linux. If you do pass --datadir, it must be writable: an unwritable one aborts in set_data_dir rather than segfaulting, which is a different failure with the same outcome.

The workspace guard covers tool arguments, not the scripts themselves. OpenSCAD's include, use and import can reach any file the process can read, so running an untrusted .scad is running untrusted code.

License

MIT. See LICENSE.

Not affiliated with Creality, Bambu Lab, Elegoo, the OrcaSlicer project or the OpenSCAD project. Those names appear here only to describe what this drives.