Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/jjh111/MetaMedium itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 21, 2026dyna.ink
A recombinatorial drawing system — interfaces that learn your visual vocabulary, recognize compositional patterns as you draw, and use AI as a "meta-word" interpreting over geometrically grounded strokes.
dyna.ink was formerly MetaMedium: the repository and the addresses below still carry that name, and RENAME-PLAN.md says what changes and when.
Start Here
📄 Interactive whitepaper (v5): dyna.ink: AI Beyond Chat
Demos
| Demo | What it shows |
|---|---|
| Reference surface — the demo | The engine's reference surface: draw → circle it → check ✓ → name it → it's recognized when drawn again. Living artifacts from the drawing, handwriting read when asked, models as participants. Works offline; touch + mouse. |
| Earlier prototype (2025): Canvas demo | Heuristic recognition, spatial graph, visual vocabulary library, live "why" reasoning. |
| Earlier prototype (2025): MetaDoodle | Tiered AI interpretation: heuristics → in-browser WebLLM (WebGPU) → LM Studio local API. Plus voice input. |
| v2 PoC: drawing-responsive text | Whitepaper text reflows in real time around shapes you draw — the medium-blending thesis, live. |
| Micro demos | Fish vocabulary, composition diagrams, no-modes graph, and other small experiments. |
Earlier-prototype keyboard shortcuts: Ctrl+Z undo, Ctrl+Shift+Z redo,
Enter accept top suggestion, Escape clear selection.
Plan & Architecture
- MVP.md — the product: ink over living artifacts, and the one loop that proves it
- KEYFRAMES.md — next sprint: shape → diagram-shape → code, each a closed vocabulary
- ROADMAP.md — status, the August 2026 accounting, and what's next
- ARCHITECTURE-v7-PARTICIPANTS-AND-TIERS.md — active plan: a model in the loop, and the conversation benchmark
- ARCHITECTURE-v6-SESSION-ENGINE.md — active design: the no-modes session engine
- metamedium-core-schema.md — graph data model: everything is a node
- ARCHITECTURE-v5-UNIFIED-ENGINE.md — partly superseded; reference for the deferred MoE/embedding work
- archive/PRD-v4-LLM-Grounded.md — archived; still the spec for tiered escalation and the unbuilt MCP server
- CLAUDE.md — codebase guide and the repo's single inventory (structure, conventions)
Experiments
EXPERIMENTS.md — the side tier: lens-canvas (infinite
canvas + confidence-routed lenses), the drawing-responsive text PoC, vision/LLM
probes, and the explainer video. Experiments de-risk platform bets and feed
proven ideas back into core; they aren't the product.
Open a folder as a canvas
The reference surface (Demos/session-engine.html) treats a folder as the
canvas: every file of a known kind (html, js, json, svg, md, txt,
png, jpg) becomes an artifact you can draw on, and your ink is saved
beside them as your own log under .metamedium/logs/. Three ways in:
- A folder on disk — the folder tile in the control centre (Chrome or Edge).
- A live room — the live tile, or
session-engine.html?live=<room>: other tabs on this machine that open the same room see each other's ink as it lands, each hand in its own colour. Across machines, runnode Demos/relay.mjsand add&relay=http://<host>:8020. Claude Code is a hand too:node Demos/mcp.mjs(registered in the repo's.mcp.json) joins roomclaudeand gives Claude seven tools — look, see the ink, draw, say, propose, transcribe, write — and the live pane's with Claude button joins the same room from the canvas. Autosave rewrites only your log file; git carries it between machines. - A repository —
session-engine.html?git=owner/repo(orowner/repo@branch,owner/repo/some/dir). Reading a public repository needs no token; writing needs a fine-grained token you paste in, and the canvas holds it on your device only if you say so. Every save is a commit of your log file. This is what lets a phone draw on a repository. - A published site —
session-engine.html?folder=https://…/opens a static deploy read-only through its.metamedium/manifest.json, so anyone can draw on it and the ink stays theirs.
The page is installable (a web manifest and a service worker cache the shell), so it opens with no network; a folder's files and a model's API are always fetched live.
Development
The engine — geometry, recognition, spatial graph, and the no-modes
session engine — lives in core/ (TypeScript,
zero runtime deps):
cd core
npm install
npm test # 86 tests incl. the canonical-loop executable spec
npm run build
The React/TypeScript app lives in Web App Skeleton/:
cd "Web App Skeleton"
npm install
npm run dev # development server
npm test # vitest suite (geometry, recognition, spatial)
npm run lint # eslint
npm run build # typecheck + production build
CI runs typecheck, lint, tests, and build for both packages on every push and PR, and fails if the committed browser bundle drifts from core.
The standalone HTML demos are self-contained single files — serve the repo
root with python -m http.server 8000 and open them directly.
Previous Versions
- Whitepaper v4 (superseded by v5)
- Earlier demo iterations live in
archive/ - Supplemental resources: jhanacek.net/metamedium-resources-65
License
The code is open source under the GNU AGPL-3.0 (AGPL-3.0-only,
LICENSE), and the whitepaper's prose and figures are
CC BY 4.0. NOTICE
says who holds the copyright and lists the third-party pieces;
TRADEMARKS.md says how the names may be used, and
CONTRIBUTING.md how to take part.