fantafnx / deckThe design file agents can actually edit.
Fanta is a native macOS editor where a cursor, source file and agent operate the same structured document—with real selection, systems, history and generation built in.
4 + 5
Agent interfaces
Native tools + opt-in MCP
.fig→FNX
Editable import
Structure, not a preview
1 undo
Atomic content batch
Commit or rollback together
01 / read
design_state
02 / batch
design_edit
03 / commit
one undo
04 / inspect
screenshot
Creative work is now a chain of lossy handoffs.
Generators create pixels. Design tools manage structure. Code tools ship products. Agents sit beside all three — but own none of the document.
.png
Generation export
Recipe stays behind
.fig
Manual rebuild
Structure is recreated
chat
Decision trail
Outside the file
Prompt
creative brief
Generate
model output
Export
flattened pixels
Rebuild
manual layers
Handoff
final file + chat
Final recipient gets pixels, not the recipe, semantic structure, or decision trail.
A design editor built as an agent runtime.
Not a prompt box beside a canvas. Fanta gives the human, the file and the agent one editable model — so generated assets and manual or automated changes remain addressable and reversible.
Humans keep precision
Direct manipulation, pen tools, layout, variables, components and inspector controls.
Agents get semantics
Pages, selection, viewport, node geometry and screenshots — not a flattened approximation of the canvas.
The document remains portable
Readable FNX lives on disk and works with the rest of the project.
One state graph is rendered for the human and exposed semantically to the agent.
editor
unmetered
hosted
image GPU
own key
agent chat
The editor is the product. AI is native to it.
A live recreation of the working editor surface — rendered from current interface components, not a stale screenshot.
Canvas
GPU-presented native surface
Systems
Components, modes and tokens
Interop
Open .fig, edit as FNX
Motion
Prototype + timeline model
The agent acts on the document — not around it.
Fanta exposes the real selection and node tree, accepts a batch of typed operations, renders the result and commits the whole change as one undoable transaction.
1 txn
Atomic edit
Content ops commit together
4 tools
Native bridge
State, edit, see, place
1 diff
Review surface
Human and agent history
read
design_state
plan
typed ops
edit
design_edit
commit
one undo
inspect
design_screenshot
Canvas, source, agent. One history.
Pages and components serialize to readable FNX; companion .ids sidecars preserve identity across source edits. Canvas changes, source changes and agent edits meet in the same history.
Readable by tools and people
JSX-like design source can be inspected without flattening the visual surface.
Change is reviewable
Manual and agent edits share undo, history and a clean structured diff.
Visual and source protect each other
Invalid FNX keeps the last valid canvas; disk and canvas conflicts use three-way reconciliation.
Capabilities route by alias. The file stays put.
Product-facing capability aliases route to a runtime catalog without a client release. Hosted image generation is deployed end-to-end; local editor work stays unmetered.
1
GPU family deployed
Hosted image generation
€0
Failed-job bill
Blocked and failed jobs
$0.01
Retail per credit
One hosted ledger
0
Credits for editor
Local design stays free
Fanta removes the reconstruction tax.
The win is not “more AI.” It is preserving editability, intent and history while intelligence moves through the workflow.
Generated media stays addressable
A result lands as a self-contained bitmap layer with bounds, prompt, model and generation provenance.
Agent work stays reviewable
Typed content batches land as one undo; screenshots let the agent inspect the rendered result.
The project remains portable
FNX and assets live together instead of scattering context across apps and chats.
Brief
prompt.txt
Generate
model session
Flatten
asset.png
Rebuild
design file
Handoff
file + messages
sets intent + selection
selection.ids · constraints
reads + batches edits
node IDs · typed ops
returns placed bitmap
bounds · prompt · model · generation ID
refines + approves
layers · history
no export boundary
placed bitmap becomes a project asset
no manual rebuild
the surrounding layout stays editable
no context reset
every change shares history
Start where the pain is already expensive.
AI-first creators already stitch models, node graphs and design apps together. Fanta gives them one native loop, then expands through the same document into teams and new media.
01
Creator wedge
Fastest feedback loop
02
Team expansion
Shared systems + usage
03
Platform path
New surfaces, same core
AI-native creators
Image generation + precision editing
Brand & product teams
Systems, review and shared libraries
Motion & 3D studios
More first-class media surfaces
Agentic creative OS
Components, recipes and workflows
One substrate compounds across every segment.
Free editor. Usage-led compute revenue.
The native product earns adoption; hosted generation earns revenue. The meter follows model usage, while local creative work stays available. Prices are provisional; AI access, allowances and seats are confirmed with the invite.
canvas
FNX files
systems
history
examples
canvas · vectors · source · history
execution
runs on the creator's Mac
0 credits
subscription refill
recurring allocation
credit pack
optional top-up
job_succeeded
hosted model returned usable output
job_failed
safety-blocked, errored, or no output
An empty hosted balance stops new hosted jobs; it does not lock the editor, the FNX document, or local editing tools.
The moat is the integration depth.
Each layer makes the next one useful. Without a real editor, agent ops are a demo. Without a structured document, history is shallow. Without routing and provenance, generation is just another export.
4
Compounding layers
Built around one model
1
Shared substrate
Every new surface reuses it
Generation / compute plane
route(intent) → generation
model catalog · policy · credits · provenance
generated images land inside the file
Agent transaction layer
read → atomic batch → inspect → iterate
typed content ops · screenshot capture · one undo
agents can act safely, not just suggest
Structured document
stable nodes → FNX + .ids → history
components · variables · reconciliation · provenance
intent remains addressable and portable
Native renderer
scene graph → native render
Metal · Skia · hit-test · selection · local GPU
the shared state has a precise surface
Pre-launch. Past the prototype.
The difficult product layers already run together. The next milestone is distribution: put the native build under real creator load and turn usage into a tighter roadmap.
Private
Current build
Integrated, pre-launch
macOS
First release
Native distribution
4
Core systems
Editor, doc, agent, cloud
Native editor
canvas · systems · motion
FNX document
stable IDs · source · history
Agent bridge
state · edit · screenshot · place
Compute router
catalog · aliases · hosted ledger
Package the release
signing · notarization · update channel
Harden project lifecycle
recovery · migrations · compatibility
Make first-run legible
onboarding · sample project · guidance
Operationalize hosted jobs
reliability · settlement · support path
public macOS alpha
outcome after release gates, not current status
Make design software operable by intelligence.
We are raising to launch the macOS alpha, grow the hosted model plane and establish FNX as the editable document for agent-native design.
Fund the product, operating system, and portable document contract as one compounding program.
input is intentionally non-numeric
Launch
turn the integrated build into a product creators can adopt
Scale
make local and hosted work reliable as usage grows
Standardize
make the document and agent contract durable
Agent-native design standard
A shared document contract where people, agents, and models create through the same precise, portable state.