Brief frontier AI in minutes, not hours.
Drop your scattered thoughts as cards on a canvas. Drag them together to connect them. Struppëflo compiles the layout into the structured prompt long-horizon AI work actually needs — sections, dependencies, execution order — and runs it with Claude, ChatGPT, Gemini, MiniMax, or Kimi.
Free · no account · runs with Claude, ChatGPT, Gemini, MiniMax, or Kimi
This is the real editor. Drag a card near another to link it — the compiled prompt updates live.
Dump
Paste everything, one thought per line. Each line becomes a card on the canvas.
Arrange
Drag cards into zones and links — or let AI group and connect them for you.
Run
The board compiles to a structured prompt. Run it in-app or hand it off to Claude, ChatGPT, Gemini, MiniMax, or Kimi.
Long-horizon work fails on missing structure, not missing models.
Frontier models now spend minutes — sometimes hours — on a single task. They can write codebases, run research, plan and execute. The thing that still breaks them is soup: prompts that wander, lose thread, contradict themselves by the third section. Struppëflo gives you the structure by construction. Zones become sections. Links become dependencies. Spatial layout becomes execution order. The model gets exactly the scaffolding it needs to stay coherent when the task runs long.
<board name="GTM Plan" cards="8" sections="3">
## 1. Research
<note id="c1" title="Target audience">…</note>
## 2. Strategy
<task id="c4" title="Landing copy"
depends_on="c2" inputs="c1">…</task>
<execution_order>
1. Value prop
2. Landing copy
3. Launch post
</execution_order>Two ways to link
Drag a card close until it glows, or pull from its edge handle. Click a line to set depends-on or input-to.
Zones
Named regions on the canvas. Each one becomes a section in the compiled prompt.
Prompt X-Ray
The compiled prompt, live — Markdown or JSON, token estimate, dependency-cycle warnings.
AI organize
Twenty half-thoughts become named, grouped zones in one click. Undoable.
Spark questions
The board asks what's missing — prerequisites, success criteria. Answers land as cards.
The basics, done right
Tabs, templates, undo history, ⌘K palette, keyboard shortcuts, JSON export.
Run with the frontier model you already trust.
One board, five model families. Each provider has a tailored system prompt so the compiled board renders cleanly in its native format. Bring your own key — Struppëflo never sees it.
Building in public.
Open source under AGPL-3.0. Now / Next / Later is a snapshot — it shifts as we ship.
- ·Run with Claude, ChatGPT, Gemini, MiniMax, or Kimi
- ·Per-model system prompts and structured output
- ·Run output → board auto-split (close the loop)
- ·Thumbs feedback on run quality
- ·Opt-in structural telemetry under Help → Help improve
- ·Share-link: read-only public boards
- ·Cloud sync across devices (Pro)
- ·Hosted backend — no BYOK required (Pro)
- ·Per-model cost estimate in Prompt X-Ray
- ·Prompt diff: see what changed since last edit
- ·MCP server: boards as resources for Claude Desktop, Cursor, Windsurf
- ·Real-time collaboration on shared boards
- ·Team workspaces with SSO and seat management
- ·Templates marketplace with creator revenue share
Local-first
Boards live in this browser. Export JSON anytime; imports are validated and sanitized.
Zero backend
A static app. The security policy blocks every outbound call except the supported AI provider APIs.
Your keys, direct
AI runs on your own provider keys, browser to API. Anthropic, OpenAI, Gemini, MiniMax, or Kimi. Without a key, deterministic local heuristics.
Five pre-wired boards. Pick one and start from structure.