The file is the truth
One canonical text format. Printing a parsed document returns the bytes you started with, so diffs are real diffs and two replicas at the same version agree exactly.
Agents
Every edit attributed. Every edit undoable.
Agent-native graphics
Design at the speed of thought
A node-graph design tool with one command bus, a text format that round-trips exactly, and an agent that has no powers you don't.
The node-graph design studio
One document. Every way of seeing it.
Canvas, layers, graph and code are four projections of the same typed node graph. Select in one and it lights up in all of them — people, agents and plugins edit the exact same thing.
Layers
Filter layers…
◈tokens
▤Page 1
▣Hero
▢Content
THeadline
◆Rule
TSub
▨Product shot
Components
❖Button / primary
❖Card / feature
Hero · 1440 × 810
Design at the speed of thought
The node-graph editor your agents can drive.
832 × 180
+ Effect
Ada
Agent · fixing spacing
49%
CodeGraphConsoleTimeline
1lumina 1.0
2doc @d_root01 {
3 tokens @tk_brand {
4 accent: p3(0.88 0.32 0.18)
5 ink: #1a1a1e
6 space: scale(4 8 16 24 40 64)
7 }
8 artboard "Hero" @ab_hero w:1440
9 frame "Content" layout:(mode:flex
10 gap:24 padding:[64]) {
11 text "Headline" @tx_head
12 weight:700 size:64 fill:$ink
13 shape.rect "Rule" fill:$accent
14 fx:[fx.shadow blur:12]
15 image "Product shot"
16 fx:[ai.remove_bg]
17 }
18 }
19}InspectTokensAgent
text @tx_head
Transform
X64
Y212
W832
H180
Text
FontInter
Weight700
Size64
Appearance
Fill$ink
Opacity1
Effects
Shadow 12 · 0.25
832 × 180
Clean
Live · 3
− 49% +
Four projectionscanvas · layers · graph · code
Design PR ready−5 lint findings · $0.04
Background removedon your GPU · nothing uploaded
Built on engines you already trust
- Skia · CanvasKit
- WebGL 2
- Loro CRDT
- Yoga layout
- HarfBuzz text
- Transformers.js
- ONNX Runtime Web
- SQLite · OPFS
- Model Context Protocol
- Cloudflare Workers
Flyers, posters & social posts
From a blank page to a finished flyer
Start from one of 900+ templates, backgrounds and graphics — or describe what you need and let the agent compose it. Every layer, effect and word stays editable.
300+ templates, 150+ backgrounds, 480+ graphicsChurch, events, social, adverts, business and lower thirds, plus icons and CC0 photos. Every one is a plain
.lum file you can edit.Effects that look finished30+ effect nodes — 3D extrude, bevel, glow, bloom, rim light, halftone, grading — and one-click Styles like 3D gold, chrome and neon.
The right page, then every page outStart on a flyer, social, screen or paper size (px, mm or inches), add pages, and export them all as one PDF or a zip.
Free pictures, found for youAsk for a lighthouse, a loaf of bread, a stained-glass window. The agent searches the Creative Commons library — CC0 and public domain by default, with a credit line when one needs it — adds the picture to your assets with its licence, and places it.
Crop anything, resize without surprisesCrop a photo, a shape or a title in a mode that shows the whole picture behind the window. Handles keep a picture's proportions — hold Shift to stretch — and a group scales its contents.
Recolour in one stepSwap a group's colours for your palette, or shift its hue and saturation, and every colour keeps its transparency.
Foundations
Three decisions everything else follows from
One way in
Every mutation crosses the same bus: validated, capability-checked, invariant-checked. There is no side door for the UI, for plugins, or for the agent.
Agents earn their reach
Tools are thin veneers over that bus. An outside agent gets the scopes a human ticked on a consent screen — and merge and pixel scopes start unticked.
Four projections
Never have to see a wire. Always able to.
Start in a canvas that feels like the tools you know. The graph builds itself underneath — open it when you want the power, read it as code when you want the precision.
Design at the speed of thought
The node-graph editor your agents can drive.
Open the studioRead the plan
- ▣Hero artboard
- ▢Content frame · flex
- THeadline text
- ◆Rule shape.rect
- TSub text
- ▢Actions frame · row
- ❖Primary instance
- ❖Secondary instance
artboard "Hero" @ab_hero w:1280 h:720 fill:$paper {
frame "Content" layout:(mode:flex gap:24 padding:[96]) {
text "Headline" @tx_head weight:700 size:72
fill:$ink "Design at the speed of thought"
shape.rect "Rule" w:220 h:6 fill:$accent
text "Sub" size:20 fill:$muted "The node-graph editor…"
frame "Actions" layout:(mode:flex direction:row gap:16) {
component.instance of:@btn_primary
}
}
}Selection is shared — Headline is highlighted in every projection.
AI design agent
An agent with no powers you don't
The agent is a peer on the command bus — the same tools, the same rules, every edit attributed and undoable. You decide how far it reaches.
Read & lint
Edit live
Branch
Merge on policy
Tighten the hero and fix lint
agent/tighten-hero → main · 4 commands · 18 s
BeforeAfter
frame "Content" layout:(mode:flex
- gap:18 padding:[57 64]) {
+ gap:24 padding:[64]) {
- text "Headline" size:66 fill:#1b1a1f
+ text "Headline" size:64 letterSpacing:-0.02 fill:$ink
- shape.rect "Rule" w:220 h:6 fill:#e2542f
+ shape.rect "Rule" w:220 h:6 fill:$accent
text "Sub" size:20 fill:$muted- align/near-miss
20 - color/off-palette
20 - spacing/rhythm
10 - Lint findings
50
From a brief to a finished design
Describe it in a sentence. The agent looks up what it needs, writes the design itself, then checks its own work.
- 1AsksWhen the brief is thin — date, venue, who it is for — and takes steering while it works.
- 2Looks things upSearches the web for facts it was not given — a brand's colours, a venue's hours, service times — reads the company's own pages, and finds free pictures.
- 3ComposesLayout, type, colour and effects written as
.lum— not a template re-skin. - 4ChecksMeasures that text fits, reviews overlaps and contrast, and fixes what the linter finds.
- 5Hands overEvery step attributed and undoable. Tag
@agentin a comment and it answers in the thread.
Explore colour variants
Four real branches, each rendered headlessly and scored on lint, composition and brand fit.
The platform
Photoshop depth. Canva ease.
Everything is a node — so everything is editable, diffable, undoable and scriptable.
On-device AI — or in the cloud
Remove backgrounds, segment, upscale and estimate depth on your own GPU — as re-runnable nodes, not destructive filters — with every model licence-audited for commercial use. Rather not download one? Signed in, background removal and picture reading can run in our cloud instead (processed, not kept). Set “This device only” and image AI never leaves your device.
Local-first collaboration
A CRDT document that works offline and merges cleanly. Presence, comments anchored to nodes, follow mode — and tag @agent in a comment to hand it a task.
Lint that fixes itself
Every finding carries its own fix as an executable command patch. “Fix all” converges.
Design system as data
Colour, type and spacing tokens that every layer references — and every linter, agent and pipeline respects.
Auto-layout, constraints, breakpoints
Flex and grid on a real layout engine, expression constraints, and breakpoints that are sizes of one design — Mobile, Tablet, Desktop — each remembering its own tweaks.
Motion on any property
Keyframe any number or colour, scrub it on the canvas, export GIF or MP4.
Import everything. Export anything.
Figma components become definitions and styles become tokens; PSD layers, masks and blend modes come across as nodes. Export every page as one PDF, or a zip of PNGs or SVGs.
Branches, checkpoints, history
Filter history by actor — “everything the agent did yesterday” is one click.
Plugins & skills
15 official plugins — QR codes, charts, schedules, tickets, scripture, photo grids, magic resize, bulk create and more — plus 5 built-in skills. Record a macro once and your agent can run it as a skill. Plugins are sandboxed, with declared capabilities.
MCP server for design
Opens in a terminal over MCP
The same engine runs in a browser tab and in Node. Point any MCP client at the bridge and your agent reads, edits, lints, renders and exports documents — under the scopes a person approved.
- 49 MCP toolsPerception, action, branching, graphics, free pictures and plugins — the in-app agent's own tools (bar its web search, which your agent brings itself), plus session and pairing.
- Device-code pairingApproval can narrow the requested scopes; it can never widen them.
- A scriptable studioThe
lumina.*console is the client the agent uses. Save a script as a skill. - An open, deterministic formatUnknown node types round-trip untouched. Diff it, review it, generate it.
$ claude mcp add --transport http lumina https://<your-bridge>/mcp
✓ Connected · 49 tools
› tighten the hero and fix what the linter finds
pair_start → approve KXWD-7Q3M in the studio
✓ granted doc.read doc.edit lint.run (branch.merge withheld)
• doc.read outline · 14 nodes
• lint.run 5 findings
• edit.dsl @tx_head letterSpacing:-0.02
• doc.apply 4 commands on agent/tighten-hero
✓ lint.run 0 findings · design PR ready for reviewlumina 1.0
doc "Launch" @d_launch {
tokens @tk_brand {
accent: #b8380c
ink: #16161a
type: scale(14 16 20 28 40 72)
}
page "Site" {
artboard "Hero" w:1280 h:720 fill:$paper {
text "Headline" weight:700 size:72 fill:$ink
"Design at the speed of thought"
}
}
}// The console is the same client the agent uses.
const heads = lumina.query("text[name~=Headline]")
await lumina.transact(
heads.map((id) => ({ op: "node.set", id, props: { weight: 700 } })),
"bolder headlines",
)
const findings = await lumina.lint()
return findings.map((f) => f.rule)name: Design CI
on:
pull_request:
paths: ["**/*.lum"]
jobs:
design:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- run: pnpm install --frozen-lockfile
- run: pnpm --filter @lumina/design-ci build
# canonical diff · lint delta · brand verdict · renders
- run: node tools/design-ci/dist/main.mjs
--base origin/main --out design-ci-outDesign CI
Design CI, because the document is code
The linter's findings carry their own fixes as executable command patches. Run them in the editor, or run them in a pipeline against a pull request — same bus, same result.
- 01Parsethe
.lumin the pull request - 02Lintagainst the workspace brand policy
- 03Renderevery artboard headlessly
- 04Gateblock the merge, or post the diff
Lumina Design CI
4 artboards rendered · 0 errors · brand policy: pass · canonical: yes
Engine
Built like a game engine
Skia in a render worker, a tiled compositor, and one contract: interaction never waits for evaluation.
5,000+nodes per document, measured in CI
1.4 msdraw p95 at 5,000 nodes
≈2 framespointer to pixels while dragging
1 switchdecides whether the agent's model ever receives a render of your design
Frame time while dragging · 5,009-node document
16.7 ms budget
Enterprise
Governance your agents can't argue with
The same policy gate runs in the studio, headless, and in CI. Identity follows your directory, every action is attributable, and your model traffic can run on your own keys.
SSO & SCIMOkta, Entra and Google via WorkOS. Roles follow groups; deprovisioning revokes tokens immediately.
Brand policy enforcementAdvise, warn, block export or block merge — identical in-app, headless and in CI.
Tamper-evident auditHash-chained records you can re-verify, attributable to actor, task and approver.
Scoped agent accessHuman-approved device codes. Merge and pixel scopes start unticked.
Bring your own model keysRoute agent calls through your provider account and compliance boundary.
You decide where pixels goOne switch — “Let the model see the canvas” — decides whether the agent's model receives a render of your design. “This device only” keeps image AI off our servers.
Architecture
Seven layers. One direction.
A DOM-free core with dependencies that only point down — enforced in CI. The same engine renders in a browser worker, in Node for your pipelines, and in the MCP bridge for your agents.
- schema → dsl
- A typed node taxonomy and a canonical, byte-stable text projection.
- doc → commands
- A movable-tree CRDT behind one bus: validation, capabilities, invariants.
- layout · eval → render · lint
- Yoga layout, pure evaluation, a Skia tile compositor, fix-carrying lint.
- agent
- Orchestration over the same tools — no private API, no special powers.
Pricing
Free where it costs us nothing
The editor and every on-device model run on your hardware, so they are free forever. Pay for the cloud, the agent and the governance.
Free
For individuals and local work
$0forever
Open the studio- Full editor, unlimited local documents
- All on-device AI, free forever
- 25 agent credits a month, design PRs included
- 3 shared documents
- 20 cloud image-AI runs a day
- 50 web searches and pages a day for the agent
Pro
For designers who work with agents
$16per month
Start with Pro- Everything in Free
- 900 agent credits a month
- Unlimited shared documents
- 200 cloud image-AI runs a day
- 300 web searches and pages a day for the agent
Team
For design teams and their agents
$28per seat / month
Start with Team- Everything in Pro, with 900 agent credits per seat
- Unlimited members, with owner, editor and viewer roles
- Brand policy that advises and warns
- Publish plugins & skills to the marketplace
- 500 cloud image-AI runs a day
- 1,000 web searches and pages a day for the agent
Enterprise
For organisations with policy to enforce
Customannual
Talk to us- SSO & SCIM, audit log
- Policy enforcement to block export or merge
- Bring your own model keys
- Private MCP bridge, Design CI, self-hosted relay
- Unmetered cloud image AI and web research
Upgrade from inside the studio — pay by card or PayPal, or with mobile money (MTN MoMo, Orange Money). Agent credits renew every month.
FAQ
Questions, answered plainly
Is Lumina really free?
Yes. The full editor, unlimited local documents and every on-device model are free forever — they run on your machine, so they cost us nothing to give you. Paid plans add more agent credits, unlimited shared documents and team governance.
What can I make with it?
Flyers, posters, social posts, banners, lower thirds, A4 prints, UI mock-ups and web pages. Start from 900+ templates, backgrounds and graphics, or from a blank page in the size you choose — flyer, social, screen or paper, in pixels, millimetres or inches — and let the agent compose it from a brief if you like.
Do my designs leave my device?
Documents live in your browser's private storage and sync only when you share them. When you use the agent, its model receives your prompt and the document's text — and, by default, a render of your design so it can judge the result by eye. Turn off “Let the model see the canvas” and it gets text and a description from Lumina's own image AI instead. Image AI runs on your device when it has the model; if you are signed in and have not downloaded it, Automatic mode runs background removal and picture reading on our servers — processed, not kept — and AI Retouch always runs there. Choose “This device only” and image AI never leaves your device.
Can the agent look things up online?
Yes, and it is on by default. Give it a brand or a venue and it can search the web, read the company's own pages and use what they say — colours, slogans, opening hours, service times — instead of guessing. It can also find free pictures in the Creative Commons library and add them to your assets with their licence. Web research runs on our servers, so you need to be signed in, and each plan has a daily allowance (see Pricing). The words it searches for go to a search service or to the Creative Commons library; the agent is told to keep your design's private content out of them. Switch either off in Preferences → AI & agent.
What is a .lum file?
The canonical text form of a Lumina document. Printing a parsed file returns exactly the bytes you started with, so version control, code review and diffs work the way they do for code — and node types a reader does not know round-trip untouched.
How do external agents get access?
Through a device code that a person approves in the studio. Approval can narrow the scopes an agent asked for, never widen them, and sensitive scopes such as merging or pixel access start unticked.
Which AI models run on-device?
Background removal (ormbg), segmentation (SlimSAM), upscaling (Swin2SR), depth estimation (Depth-Anything-V2-Small) and picture reading (Florence-2), through Transformers.js on WebGPU with a WebAssembly fallback. Each is a one-time download of 40–340 MB, and only after you say yes. Every on-device model is audited for commercial use — MIT, Apache-2.0 or BSD only.
Does it work on a phone?
Yes, in your phone's browser. The studio shows one panel at a time with a bottom bar — Canvas, Layers, Agent, Design, Graphics, More — and signed in, image AI can run in the cloud so there is no model to download over mobile data.
Can I import from Figma or Photoshop?
Yes. Figma files import over the REST API with components as definitions and styles as tokens; PSD files bring layers, groups, masks and blend modes across as nodes. Export to PNG, SVG, PDF, animated GIF or MP4.
Can we self-host?
The core runs as static files. The optional services — sync relay, accounts, AI proxy, marketplace and the MCP bridge — are Workers and a Node container you can deploy in your own cloud account.
Local-first. Text-first. Yours to drive.
Runs in the browser with no server. Opens in a terminal over MCP. Same document, same rules, either way.