yuzu-block (working title)

An open-source public site and auth-free MCP server that render anonymous, aggregate-only data from a voice journal built on ElevenLabs Scribe transcription as a 3D block sculpture, maintained by STUDIO CON. This repo is independent from the voice journal's own codebase; it consumes aggregate numbers only.

What it shows / what it never shows

Shows weekly aggregate rings:

• record count → block height

• silence-day ratio → gaps in the ring

• words-per-record → where a week sits on the ink ramp

The sculpture itself carries no text, numbers, or labels. Page chrome around it is limited to the YUZU logo (linking to yuzu.style), a lead, and a footer; every string on the page is defined in lib/copy.ts. No metrics — record counts, per-user data, consecutive-day counts — are displayed anywhere. No counters, no congratulatory effect, no light or sound cue. The sculpture never exceeds the server snapshot it started from — see "Rendering" for what it does show while a visitor watches. Weeks whose cohort is below the anonymity threshold render as empty rings.

Related MCP server: oura-mcp

Concept

Open the material, never the interpretation.

Annual rings: layers are cut by time passing, not by effort. Numbers are counted, not praised.

Setup

npm install
npm run dev

Open /. The MCP endpoint is at /api/mcp.

npm run check
npm run build

Requires Node 20+.

MCP server

No authentication. Streamable HTTP at /api/mcp. One tool: get_ring_data.

Input:

{ "year": 2026 }

Trimmed example output (RingAggregate, buckets truncated):

{
  "year": 2026,
  "unit": "week",
  "minCohort": 50,
  "generatedAt": "2026-09-12T00:00:00.000Z",
  "buckets": [
    {
      "index": 0,
      "start": "2025-12-29",
      "sufficient": true,
      "recordCount": 214,
      "wordCount": 18904,
      "silenceDayRatio": 0.412
    },
    {
      "index": 1,
      "start": "2026-01-05",
      "sufficient": false,
      "recordCount": null,
      "wordCount": null,
      "silenceDayRatio": null
    },
    {
      "index": 2,
      "start": "2026-01-12",
      "sufficient": true,
      "recordCount": 233,
      "wordCount": 21172,
      "silenceDayRatio": 0.388
    }
  ]
}

Custom connector

Add as a custom connector using the URL only:

https://<your-deployment>/api/mcp

curl example

Initialize:

curl -s http://localhost:3000/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2024-11-05",
      "capabilities": {},
      "clientInfo": { "name": "curl", "version": "0.0.1" }
    }
  }'

Call the tool:

curl -s http://localhost:3000/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/call",
    "params": { "name": "get_ring_data", "arguments": { "year": 2026 } }
  }'

Data contract

export interface WeekBucket {
  index: number;
  /** ISO date, YYYY-MM-DD, Monday of the bucket's week. */
  start: string;
  /** False when the cohort for this bucket is below the anonymity threshold. */
  sufficient: boolean;
  recordCount: number | null;
  wordCount: number | null;
  /** 0..1 — days with zero records / days in the bucket. */
  silenceDayRatio: number | null;
}

export interface RingAggregate {
  year: number;
  unit: "week";
  minCohort: number;
  generatedAt: string;
  buckets: WeekBucket[];
}

Buckets are ISO weeks: Monday-start, week 1 is the week containing January 4, and a year has 52 or 53 of them. The reference upstream computes these boundaries in Asia/Tokyo.

Anonymization policy: week-level granularity only. Buckets whose cohort falls below minCohort are returned with every numeric field set to null and sufficient: false. No per-user field exists anywhere in this shape, by construction.

The data source is selected by the DATA_SOURCE_URL environment variable (lib/data-source.ts):

• Unset: a deterministic mock (lib/mock-aggregate.ts), seeded by year and date.

• Set (e.g. DATA_SOURCE_URL=https://app.yuzu.style/api/public/weekly): GET {DATA_SOURCE_URL}?year=YYYY, expected to return JSON matching RingAggregate exactly (52 or 53 WeekBuckets, unit: "week", nulls agreeing with sufficient). The response is validated against that shape; any failure — network error, non-2xx, timeout, or invalid shape — falls back to the mock and logs one line server-side. The upstream is expected to set long CDN cache headers; no auth is sent or required.

When a validated upstream payload has no week above the anonymity threshold at all, the site and MCP tool fall back to the mock as well and report source: "mock".

The MCP tool's JSON output carries a top-level source field ("upstream" or "mock") alongside the aggregate, so a caller can see which one produced a given response. The page never renders this field.

wordCount's unit is defined by whichever source produced it. For the reference upstream (the voice journal above), it is transcribed characters, not words.

Rendering

The scene is a snapshot: SceneSpec is generated server-side on every request and handed to the client once. On top of it the client holds back about 10% of the blocks at first paint and reveals them one at a time on a bounded schedule, and independently slides whole cell columns along the ink ramp while holding the mean tone close to the snapshot's. The sculpture reads as continuously carved, but it converges to the snapshot and never adds anything beyond it.

prefers-reduced-motion renders the finished snapshot with no camera motion and no carving, and draws on demand rather than running a render loop over a still image.

Camera

Orthographic, axonometric. It turns continuously and also eases its elevation between 31 and 53 degrees over about a minute — yaw alone reads flat, because the face angles never change.

The frame covers the viewport rather than fitting inside it: the solid runs off all four edges and under the page chrome, which the lead copy sits on top of by design. Nothing is reserved for the chrome.

A square footprint is sqrt(2) narrower face-on than corner-on, so the frustum is re-fitted to the live orbit angles every frame and then pulled partway back toward the widest case (dampedExtents). Fitting the live silhouette alone swings the framing by the full sqrt(2), which crops past the point where the solid still reads as a solid.

Surface

Unlit. There is no lighting in the scene, so the six per-face constants are not brightness — they are ink coverage. Two independent reads of one ordered (Bayer) screen decide every pixel:

• how much ink, from the face's coverage, against the stock beneath it

• which ink, from the block's position on the ramp, between the two stops it falls between

Both are hard thresholds, never a blend, so every pixel lands exactly on a palette token and no in-between colour is produced. Anti-aliased silhouette edges are the one exception. The screen is anchored to the solid rather than to the viewport: anchored to the viewport the faces slide across a fixed grid of dots as the solid turns, and the whole surface scintillates.

The ramp is in lib/palette.ts as INK_RAMP, light to muted:

YUZU_PALE #FCEE8A · YUZU_LEMON #F8E262 · YUZU_YELLOW #F5D84A · YUZU_GOLD #EBCB4A · YUZU_STRAW #E2C652 · YUZU_LINEN #CCBB6B · YUZU_STONE #BEB47D · YUZU_ASH #B3AC8E

SURFACE_BORDER #E8E0C8 is the stock the ink sits on, and YUZU_WHITE #FAFAF5 is the page — so the only true white inside the solid is a real void, which is a cell with no block, which is silent days.

A block's place on the ramp comes from a smooth field across the footprint, per-column noise and per-block noise, plus the week's own measure when there is real data. Blocks are ranked against each other and the ramp's uneven stop positions divide the surface between the inks; see lib/tone.ts.

Ground

One plane under the solid carries the nominal 52-week ring grid as concentric squares — most of that grid emits no blocks, so this is the stock the sculpture was cut from — plus a slowly drifting paper screen, and a misregistration slip every 7 to 19 seconds. It is transparent everywhere it has no mark, so the page's own background shows through.

Working on this

npm run check (typecheck, lint, test, brand words) must pass before anything is considered done.

.claude/lessons/ holds findings from earlier sessions that are not recoverable from the code or the log — rendering traps, shader traps, how to read a deploy's state, test hygiene. Read it before starting.

npm run deploy:status reports what Vercel is actually doing with a commit. GitHub returns pending for a commit with no checks at all, so state alone cannot tell a dropped push from a running build; the script separates them and exits 2 when the push was never received. npm run deploy:wait polls until it settles.

scripts/colour-audit.browser.js counts the rendered ink shares by family and their spread across the frame. It has to run in the browser: what a viewer sees is weighted by which faces are turned toward the camera, not by how many blocks carry a tone.

ElevenLabs

The words that become blocks are transcribed by ElevenLabs Scribe in the upstream journal. This repo is intended for submission to the ElevenLabs Grant / Showcase as a standalone, free, open-source demo.

Deploy

Vercel, zero configuration (next build). To use a real upstream instead of the mock, set DATA_SOURCE_URL in the Vercel project's environment variables (see .env.local.example).

License

MIT © 2026 STUDIO CON.