Showcase · Kestrel
Seven plain-language briefs, one agent, Polykrom's MCP tools and nothing else. Kestrel is a desktop pick-and-place arm: a gear-driven turntable, shoulder and elbow, a coupled parallel gripper, standard hardware from the library, a motion study, a stress study and a robot description for simulators. Every number below was re-measured on the saved design by an independent script. The last section lists what this version can't do yet.
Live: the URDF Polykrom exported (one mesh per part, joints in metres) driven by the motion study's recorded joint values. Drag to orbit, scroll to zoom.
Each stage was one request to Claude Code running headless with Polykrom's MCP server as its only tool. It had no file system, no shell and no help. After each stage, a separate script opened the saved .polykrom file and replayed drives, sweeps and measurements to confirm the claims.
A Ø180 aluminium base with a bearing seat, a NEMA 17 motor, a 20 mm pinion and a 100 mm turntable gear, coupled 5:1.
A yoke on the turntable, a 140 mm upper arm with lightening holes, a 120 mm forearm, and steel pins in 8.2 mm bores.
A PLA wrist with two nylon jaws on sliders, coupled by a motion link.
A 608 bearing, four ISO 4762 M4×12 screws and four ISO 7089 washers from the library, with real materials on every body.
A 120-step study: swing 90° through the gear train, reach down, grip, lift, return, release.
Static stress on the upper arm, held at the shoulder, with 40 N square to it at the elbow.
Lighten the upper arm while keeping SF ≥ 3 and deflection < 0.2 mm, without touching its fits.
The briefs were written quickly, on purpose, the way people actually ask. The agent built what was asked, measured it, and reported where the brief couldn't work. It never quietly changed the design to hide a problem.
A 42.3 mm NEMA 17 centred 60 mm out reaches 81.15 mm, past the 66 mm inside wall of a Ø140 base. It found the overlap (6,194 mm³) and offered three fixes. We enlarged the base.
A 7 mm pocket in a 4 mm top is a hole, not a seat. We added a hub under it.
8 mm deep in an 8 mm disc. They became 6 mm, leaving 5.2 mm of thread engaged for the M4 screws.
Half of each yoke-plate hole sat under a cheek, so no screw head could seat. It shortened the cheeks from 40 to 28 mm, checked that the arm still clears them, and explained the alternative.
One small robot touches most of the app.
The brief asked for SF ≥ 3 and less than 0.2 mm deflection, with 40 N square to the arm at the elbow, its worst case. The first result was SF 43. So in stage 7 the agent was asked to make it as light as it reasonably could without touching any fit. It ran the solver after every change.
| # | Change | Mass (g) | Safety factor | Deflection (mm) |
|---|---|---|---|---|
| 0 | As built in stage 2 | 181.8 | 43.4 | 0.034 |
| 1 | Pockets from both sides: 4 mm flanges and web | 119.1 | 46.0 | 0.042 |
| 2 | Flanges and web 2.5 mm | 97.3 | 42.0 | 0.054 |
| 3 | Mid-section thinned to 12 mm | 85.2 | 19.2 | 0.075 |
| 4 | Mid-section 8 mm | 79.2 | 11.1 | 0.097 |
| 5 | Mid-section 6 mm, five Ø14 web holes | 73.3 | 10.5 | 0.129 |
| 6 | R1 fillets at the pocket steps (final) | 73.3 | 10.7 | 0.128 |
Linear static FEA through Nightfall (Netgen mesh, MFEM solve). The final run was converged over three meshes down to 0.89 mm: deflection 0.1281 / 0.1283 / 0.1284 mm. The fillets were added because the peak stress kept rising with mesh refinement at the sharp steps. An independent re-run of the saved part gave SF 10.75, 25.7 MPa peak and 0.128 mm. The fits stayed exactly as they were: Ø8.200 holes at 140.000 mm, R12 ends, full 20 mm thickness at the pins.
The animation the agent exported with motion.export_animation: 121 frames from the headless renderer.
Found by building Kestrel, not by guessing. Several were fixed the same day; the rest are on the roadmap.
A joint sweep can't turn a pin in a hole with 0.1 mm of clearance. It runs out of its collision budget, even for that one pair.
That's exactly the fit a sweep exists to check. The agent swept the other part pairs one at a time and checked the pins statically instead. → #25
Checking a 21-body study for collisions, all pairs at once, stops at step 0.
The work budget runs out, and any part that overlaps from the start (a thread in a tapped hole) eats it. The agent split the check into groups and sampled arm-against-base every 4th step. Pairs that touch at the start should be set aside, as joint sweeps already do. → #25
Joint limits are per joint, and nothing re-checks them when the assembly changes.
Each limit is safe on its own, but shoulder +114.5° with elbow −77.4° drives the forearm 2,661 mm³ into the yoke. There are no coupled limits or reachable-workspace check yet.
The elbow's −77.4° limit was swept in stage 2, before the gripper existed. Once the gripper was on, the wrist hit the upper arm at −67.7°. Polykrom didn't flag the stale limit; the agent found it in stage 7. It is the one value we changed by hand after the build. → #25
The kernel refused 1 mm fillets along the upper arm's web-to-flange edges, so those stay sharp.
Contact checks that run on meshes also raised a false 0.022 mm "contact" on a pin with a true 0.100 mm gap, from the mesh's 0.16 mm chord tolerance. The exact check was right. → #25
Library parts arrive with a placeholder material and bodies named Body1–9.
The 608 bearing seated facing out of its pocket and had to be flipped. It is a solid envelope, so in steel it weighs 18 g against a real bearing's 12 g. The screws seated on the plate rather than on their washers and were raised 0.8 mm by hand. → #43
The URDF has no <mimic> tags.
The gear between pinion and turntable, and the link between the jaws, are lost in simulators. Effort and velocity limits fall back to defaults because parts don't carry actuator specs yet. → #27
Stress comes back as numbers, not colours on the part.
Nightfall now writes the surface field. Drawing it on the body is in progress. → #29
Gears are pitch-diameter blanks, not teeth.
The kinematics are exact, but there's no tooth contact or backlash to check. → #26
The exported animation is flat grey.
The headless renderer ignores material appearance, so the 3D view above is the better picture. → #33
Building the earlier test assemblies found three bugs, all fixed:
→ #17
The design, the robot description, the bill of materials and the agent's own reports.