#Dogfooding the MCP server

To see what designing with tscad is like for an AI agent, five models were designed from plain-English briefs. The designer acted purely as an MCP client: every step was a JSON-RPC tools/call through runMcpServer over in-memory streams, using only tscad_docs, tscad_evaluate, tscad_render, tscad_check, tscad_pick_selectors and tscad_export, with no reading of src/. This page summarises each session and lists every point of friction, with what was fixed and what still needs work.

The models are in examples/agent/.

#The sessions

#1. Desk headphone hook

"a desk headphone hook that clamps to a 25 mm desk edge"

<img src="images/agent/headphone-hook.png" width="560" alt="Headphone hook">

1. tscad_docs read the guide, then the exact props of extrude, polygon and fillet.
2. Designed the side profile as one <polygon>: a C clamp over the desk edge, plus an arm with a lip below it. Extruded it on plane="XZ" with center, filleted "|Y", and cut an M6 TapHole through the lower jaw for a clamp screw.
3. tscad_evaluate succeeded first time: 27.9 cm³, 35 g of PETG. tscad_render with view: "front" showed the profile.
4. tscad_check reported 1251 mm² of overhangs, most of it the underside of the top jaw.
   • Friction: it didn't say which way up to print it. I had to guess buildDir and re-run the check.
   • Now: the check adds "print it with buildDir [0, 0, -1] (upside down): overhangs drop to 0 mm²".
5. tscad_export wrote the STEP file (61 kB).

#2. Drawer organiser

"a parametric drawer organiser for a 300×200 mm drawer"

<img src="images/agent/drawer-organiser.png" width="560" alt="Drawer organiser">

1. Built the outline with roundedRect and cut all the compartments in a single <extrude>: a grid of roundedRects, mapped into one cut, which the guide recommends for speed. Evaluated in 1.6 s.
2. The full tray is 298 × 198 mm, too big for a 220 mm bed. A piece choice param cuts it into quarters with two nested <split>s.
   • Friction: split's plane prop was typed PlaneSpec, which wasn't explained anywhere. I had to infer that "XZ" has its normal on +Y to pick keep.
   • Now: prop docs expand type names inline (PlaneSpec = "XY" (normal +Z), …).
3. tscad_check on piece: "front-left" passed, with a thinnest wall of 0.8 mm where the cut halved a divider. That's a design note rather than a tooling problem: cut between compartments.

#3. Raspberry Pi 5 case

"a Raspberry Pi 5 case with vents and a GPIO slot"

<img src="images/agent/pi5-case.png" width="560" alt="Pi 5 case">

1. tscad_docs topic: "parts" listed mountPatterns.raspberryPi4.
   • Friction: it gave no numbers (hole spacing, board size, where the board sits relative to the holes) and didn't say whether the Pi 5 matches. I had to work out outlineOffset + outline / 2 to centre the board.
   • Now: the parts docs include a table of every pattern (hole span, screw, outline, offset, pilot) and spell out the formula. There's also a raspberryPi5 pattern.
2. Base: shell walls, port cut-outs, and <BoltPattern> of <Standoff size="M2.5">. Lid: vent slots from a <linearPattern> and a GPIO slot.
   • The first version put the base/lid split below the USB stack. Comparing the numbers in tscad_evaluate (bboxes) against the port heights caught it.
   • Friction: connector positions (USB-C, micro-HDMI, GPIO) aren't in the library, so I used published offsets from memory. See "Open" below.
3. tscad_check with exploded: false found no interference. The lid's 4556 mm² ceiling is an overhang, and the check now suggests printing it upside down.

#4. Replacement knob for a 6 mm D-shaft

"a replacement knob for a 6 mm D-shaft"

<img src="images/agent/d-shaft-knob.png" width="560" alt="D-shaft knob">

1. Friction: there are no 2D booleans. A D profile (a circle cut by a flat) and flutes around a rim can't be drawn as sketch loops, because loops only nest (holes) or union.
   • Workaround: in 3D, an <intersection> of a cylinder and a box for the D bore, and a <polarPattern> of 24 cylinder cuts for the flutes.
   • Cost: 4.4 s, the slowest of the five.
   • Now: the guide says so up front and shows this pattern. A real fix needs 2D booleans (see "Open").
2. tscad_render with view: "bottom" confirmed the D bore.
3. tscad_check flagged 0.76 mm next to the pointer groove.
   • Friction: it printed "0 % of the surface is thinner than 0.8 mm", which reads like a contradiction.
   • Now: it prints "< 0.1 %".

#5. Laser-cut phone stand

"a laser-cut phone stand"

<img src="images/agent/laser-phone-stand.png" width="560" alt="Laser-cut phone stand">

1. Two side profiles, each with two mortises, plus a shelf and a back rest whose tenons pass through the mortises. Each shape is a component. <drawing layer="cut"> wraps it in <offset distance={kerf / 2} round={false}> for the cutting file. <part>s extrude the same components for an assembled preview.
2. tscad_check found two interferences of 27 mm³ each, at y 42..43.5: the crossbars were off-centre by half a sheet. Fixed, and the check came back clean.
   • Good: the interference report, with volume and location, pointed straight at the bug.
3. tscad_export format: "dxf" wrote the file.
   • Friction: the result said "parts: left, right, shelf-3d, rest-3d", the 3D preview parts, when the DXF actually holds the four drawings.
   • Now: it says "drawings: side-left, side-right, shelf, rest".

#Probing common mistakes

Typical slips went through tscad_evaluate to see how the errors read:

#Friction list

Fixed in this pass (agent guide, docs generation, error hints, MCP tool outputs):

1. The agent guide had gaps.
   • Added sections: assemblies and mates (with an example), a pointer to constraint sketches, printing (orientation, teardrop holes) and common mistakes.
   • It now covers <hull>, <drawing layer>, the kerf idiom with <offset>, and the <FingerJointBox> / <LivingHinge> generators.
   • It says up front that there are no 2D booleans and shows the 3D workaround.
2. The element index was one flat list, starting with 20 constraint-sketch elements. It's now grouped into modelling, constraint sketch, and parts library (with the import line).
3. Type names in prop docs weren't explained. PlaneSpec, CenterSpec, SelectorLike, MetricSize, MaterialName, HoleFit, MountPattern, Path3, Vec3 and Axis are now spelled out inline.
4. The mount patterns had no numbers. There's now a generated table and a raspberryPi5 pattern.
5. tscad_check didn't suggest an orientation. It now tries the other five build directions and reports betterOrientation when one halves the overhang area or better.
6. No selector isolated some faces and edges. When tscad_pick_selectors has no geometric selector that picks the item alone, it now offers an edges.within(...) / faces.within(...) box around it, ahead of the fragile index.
7. The tscad_export DXF/SVG result named the 3D parts. It now names the drawings it wrote.
8. Errors on an element the model component returns directly had no location. That's because JavaScriptCore drops the component's frame on a tail call.
   • The message now includes the element, e.g. (in <cylinder r={2} h={5} rotate={90}>).
   • The MCP server then finds it in the source and adds at model.tsx:L:C with a code frame.
9. Fix-it hints for rotate={n}, two-number box sizes and center strings.
10. Unknown lowercase tags that name a parts-library component now point to the import.
11. ReferenceErrors for tscad and tscad/parts exports now say what to import.
12. Percentages: "< 0.1 %" instead of "0 %" for a small but non-zero thin area.

Open: needs other owners.

• ~~Sketch: 2D booleans~~ Done since: <difference2d> / <union2d> / <intersection2d> (exact, fuzzed). The knob now builds its fluted rim and D bore in 2D: 4.4 s → 2.4 s (0.8 s without the top chamfer).
• Kernel:
  • <fillet radius={8}> on a 10 mm cube returns ok (962 mm³) instead of refusing a radius larger than the faces allow.
  • 24 small cylinder cuts take about 4 s.
• ~~Parts, as a backlog: board models with connector frames~~ Done since: boards.* holds every connector as data, <Board> anchors them, and <BoardCase> cuts the ports from that data. examples/pi5-case-auto.tsx replaces the hand-placed Pi 5 case; its port openings line up with the connector anchors to within 0.01 mm.
• Bun upstream: JSX source information (jsxDEV gets undefined) would give every element an exact location, including ones returned directly by a component.
• Render: the default light background with white or natural PLA is low-contrast. Agents may want background: "dark"; worth considering as the default for tscad_render.

#Verdict

All five briefs became working, checked, exported models in 2–4 evaluate cycles each. The biggest wins for an agent:

• errors that say how to fix them, at a line;
• a check that says which way up to print;
• interference reports with a location.

The tests in tests/mcp.dogfood.test.ts keep these fixes from regressing.