Nestbolt sheet nesting MCP server

Nestbolt

8.5% less material at the median than the best of Deepnest, SVGnest and Sparrow at equal 60-second budgets, and 9.6% at 300 seconds, in the final benchmark run (fresh seeds, ten held-out jobs). Sparrow still wins 8 of the 56 jobs at 60 seconds (details, every number).

Free, open-source nesting for people who cut fixed sheets: laser cutters, CNC routers, plasma, waterjet, fabric and leather. Nestbolt arranges parts so they use the least material. It runs in your browser (nothing is uploaded), on the command line, and as an MCP server that AI agents can call.

451 parts on one 850 × 425 mm sheet in 120 seconds on one CPU core: DejaVu Sans Bold glyphs in four sizes with their exact Bézier outlines, 90° rotations and 1 mm spacing. 29 letters sit inside the holes of others, and the unused material stays as one clean offcut along the top. The layout passed the independent validator. Rebuild it with python scripts/make_showcase.py --nest 120.

How it compares

Final acceptance run (2026-10-13): 75 frozen sheet jobs (35 academic, 40 maker, ten of them held out from all tuning), equal time budgets, one CPU core per tool, five fresh seeds each, every layout checked by an independent validator. The comparison is against the best of Deepnest, SVGnest and Sparrow on each job, each with the better of its two settings.

At 60 s Nestbolt uses less material than Deepnest on all 55 jobs where Deepnest produced a layout (median 9.9%) and than SVGnest on all 36 (14.4%). Sparrow, which Nestbolt builds on, wins 8 jobs. Seven are one-sheet jobs that Nestbolt hands to Sparrow itself after its own first layout, so Sparrow runs with slightly less time there; it is 0.03-2.4% better on them. The eighth (two sheets) is 0.34% better. That misses the project's own 60-second bar (wins or ties on 90% of jobs and never more than 1% worse); the 300-second, strip-parity and speed bars are met (status).

One benchmark job exactly as nested in a 60-second development run (m-mechanical-frame-inserts, seed 0, sheet turned to landscape): 138 parts, all 39 inserts placed inside frame windows, everything packed into 56% of a 2440 × 1220 mm sheet so the rest stays one offcut. The best competitor used 62% of the sheet on this job.

In 30 s Nestbolt matches the competitors' 300 s result on 10 of 11 jobs. The first final run (build 484ce8b) stopped with an error on one job; the fix is in 75cbb32 and both runs are published. Every run, including failures, is in BENCHMARKS.md; status lists the evidence per gate. The results page shows every per-job number, tool revision, seed and the hardware; python scripts/reproduce.py --phase final --final-acceptance rebuilds the datasets, reruns what is missing and regenerates it.

What works today

Shortened. The full README is on GitHub.