Flit MCP Server

Flit

Flit is a tiny little physics engine for 3js. It doesn't do much yet! but I believe!

Spheres, collisions, raycasts, gravity, and a spatial-hash broadphase — nothing more, on purpose. All simulation state lives in flat Float32Arrays (structure-of-arrays), so the CPU path doubles as the reference implementation for a future WebGPU compute backend: same buffers, same kernels, no re-architecting.

What's in the box

• World — point-sphere particles, semi-implicit Euler integration, impulse + positional-correction contact solver, infinite ground plane. Bodies that stay quiet SETTLE (per-body sleep driven by a stiction threshold), so resting piles cost ~nothing; a fast-enough impact wakes them, a slow push is held by static friction.
• Heightfield — 2.5D terrain as a grid of height posts, no meshing: bilinear height + central-difference normals sampled directly for collision, and a 2D Amanatides-Woo DDA for raycasts (index: -1).
• Raycast — Amanatides-Woo voxel walk over the broadphase grid; candidates from the 3x3x3 block around the walked path so corner clips are never missed. Differential-tested against a brute-force oracle.
• SpatialHash — uniform grid broadphase keyed by Morton (Z-order) cell codes. This is the Euclidean cousin in the LSH family: MinHash buckets documents by Jaccard similarity; this buckets positions so nearby points collide in the same bucket. Morton keys are bijective (no hash-collision pair bloat), invertible (no side table), and locality-ordered for a future sorted-array GPU backend.
• Distance & bit utilities — squared Euclidean/L1/Minkowski/L∞/Hellinger/ chi-square/KL distances (ported from FLANN), Morton (Z-order) keys, float bit-flips for radix sorting, octagonal approximate distance. See THIRD_PARTY_NOTICES.md for provenance and licenses.

For agents

Building a little three.js game or demo? Two ways in:

• Skill: skills/flit/SKILL.md — copy the skills/flit/ directory into your agent's skills path (e.g. .claude/skills/, ~/.code_puppy/skills/). It carries the 30-second integration recipe, the MCP option, measured performance envelope, and contributor rules.
• MCP server (no code needed): npm run mcp, or point your client at it:

{
  "mcpServers": {
    "flit": {
      "command": "npx",
      "args": ["vite-node", "mcp/server.ts"],
      "cwd": "<path-to-this-repo>"
    }
  }
}

Tools: flit_info, flit_reset (gravity/restitution/ground/bounds/ settleSpeed/heightfield), flit_spawn (rain/explosion/grid/fountain presets), flit_add_particles, flit_step, flit_state, flit_raycast — step/state return flat xyz positions shaped for InstancedMesh syncing plus settledCount, so you can watch a pile go quiet.

Usage

npm i flit-physics

import { World } from 'flit-physics';

const world = new World({ restitution: 0.4 }); // gravity and a floor at y=0 included

const ball = world.addParticle({ position: [0, 10, 0], radius: 0.5, mass: 1 });

// fixed timestep, e.g. from your rAF loop
world.step(1 / 60);

// sync to three.js: positions is a live Float32Array, 3 floats per particle
mesh.position.set(
  world.positions[ball * 3],
  world.positions[ball * 3 + 1],
  world.positions[ball * 3 + 2],
);

Terrain and sleep are one option each:

import { Heightfield, World } from 'flit-physics';

const terrain = new Heightfield({ rows: 64, cols: 64, cellSize: 1, heights });
const world = new World({ groundY: null, heightfield: terrain });
// bodies collide with the terrain and settle on it;
// world.raycast(...) reports terrain hits as index -1

// sleep is on by default (settleSpeed 0.01 m/s); opt out per world:
const wideAwake = new World({ settleSpeed: 0 });

Develop

npm install
npm test        # vitest
npm run build   # tsc -> dist/
npm run bench   # broadphase + full-step micro-benchmarks
npm run demo    # three.js demo scene (vite dev server)

Benchmarks

Deterministic seeds; numbers from a local dev machine, recorded at commit time (see commit messages for the full series, including rejected designs).

• bench/broadphase.bench.ts — broadphase only, N=4096: xor-hash 3.4–3.7 ms (444 pairs, ~90% collision bloat) → Morton keys with ordered probing 3.5–3.7 ms (234 pairs, exact).
• bench/world.bench.ts — full World.step, N=1024: xor-hash 0.85 ms/step → Morton ordered-probing 0.83 ms/step → 0.93 ms/step with the sequential-impulse velocity solver (4 iterations + LUT friction). Exact keys, real contacts, +11%.
• bench/scaling.bench.ts — two regimes. Fixed box (density rises): pairs scale ~N² from crowding physics. Scaled box (constant spawn density): flat O(N) ≈ 1.2 ms per 1000 through N=8000 while bodies are scattered. With gravity on, everything rains into a dense floor pile over ~2-4s (WARMUP=240 to reproduce) — and since 0.3.0 the pile SETTLES: 34.2 → 9.2 ms/step at N=8000 (3.7x, inside the 60fps budget) as bodies fall asleep bottom-up. Settled bodies are skipped by integration, planes, and mutual contacts until a >1 m/s impact wakes them. (docs/issues/001 — resolved; brief kept for the analysis and the remaining ideas: warm starting, adaptive iterations.)
• bench/raycast.bench.ts — 2000 seeded rays at N=4096: 63–68 µs/ray (vs ~160 µs for a brute-force O(N) oracle at the same density; the grid walk pulls further ahead as N grows).
• bench/morton-libs.bench.ts — codec bake-off vs npm libs (npm run bench:libs): ours 3.8 ns/encode, fast-morton MB 26.1, fast-morton LUT 43.8, @thi.ng/morton 539.9. In-house wins; the libs stay as devDependencies purely so the bake-off stays runnable.
• bench/threaded.bench.ts — WorkerWorld over a real worker thread (SAB ping-pong): kick() costs 0.014 ms on the main thread; with a 4 ms render workload a frame costs 8.84 ms pipelined vs 14.96 ms serial at N=8192 (113 fps vs 67). At N=1024 the 0.8 ms step hides entirely behind render (4.03 vs 4.80 ms).
• Rejected on measurement (see commits): 63-bit BigInt keys (13× alloc regression); sorted-array + binary-search broadphase (1.4× slower than Map probing in JS — negative result recorded).

Roadmap

• Morton-ordered broadphase cells — done, measured, shipped
• three.js demo scene — npm run demo, 220 balls in a box (?terrain for heightfield hills, ?n=8000 for load testing)
• Worker-threaded stepping — WorkerWorld: kick/waitForUpdate, SAB ping-pong zero-copy rendering
• Settled-pile solver cost — per-body sleep shipped (34.2 → 9.2 ms at N=8000); issue brief kept for the remaining ideas (warm starting, adaptive iterations)
• Heightfield terrain — shipped, no meshing; hooks left for more collider shapes
• Raycast coarse occupancy filter (skip empty space faster) if ray-heavy workloads show up — see commit 3173a6f for the analysis
• WebGPU compute backend once the CPU reference settles

License

MIT — see LICENSE. Third-party portions and their licenses are listed in THIRD_PARTY_NOTICES.md.