A reactive, typed, cellular runtime. From microcontroller to browser to mesh to cloudflare. The spreadsheet model, finally done right. The cell model, finally made real.
A cell isn't a variable. It's a value with a type, a history, an access policy, dependencies, and formulas. Cells compose. Cells sync. Cells outlive devices.
Change a value cell, watch every dependent recompute. The engine handles propagation. You write formulas; the graph does the rest.
Every cell has a history. Fork the engine like git. Rewind, replay, branch, merge. Undo without a stack.
Per-cell end-to-end encryption. The server never sees your plaintext. You choose who can read each cell.
Peer-to-peer mesh. Cells sync across devices with no central server. Offline-then-sync. Conflict-free merge.
Values have types: int, float, bool, string, list, object, blob. Formulas are checked. Bad input is rejected.
Same model in TypeScript, Rust, a 70KB browser file, a microcontroller, a visual editor, an LLM agent.
A real Quilt engine is running in this page. The cells below are live. Change a number, see formulas recompute instantly.
The cell model is the API surface. Each repo is a different runtime for the same model. Pick the one that fits your context — or compose them.
Files, programs, services. We know these. But the next era is built on cells.
A file is a static blob on a disk. A cell is a live, addressable, reactive value with a type, a history, and an access policy.
A program runs once. A graph is always running. Formulas recompute when their dependencies change. The graph is the application.
A service is a public API on a server. A room is a private shared space on a mesh. A service is one-to-many. A room is many-to-many, with an access policy.
The same cell model, but with LLMs. Each call is a cell. The cell model is uniform — values, formulas, and LLMs propagate reactively.
6 tabs of real z.ai cells: complete, code, reason, summarize, Q&A, and 3-cell chains. Each call is a cell. The graph shows propagation.
Try it →Send the same prompt to z.ai, Kimi, and DeepSeek in parallel. Compare side-by-side. Synthesize or vote (LLM judge picks the best).
Try it →Add notes, they auto-embed. Search semantically. Top-K context. LLM answers with grounded context. 6 cells, fully reactive.
Try it →4 personas, multi-step reasoning, memory as a value cell, output as a derived cell. The agent loop unfolds in the graph as it runs.
Try it →One repo. Clone it. Workers + D1 + Vectorize + KV + R2 + AI. Deploy in 5 minutes. 8 AI cell kinds built in. Real, not a demo.
Deploy →4 providers, 8 cell kinds, 6 patterns, security model, cost control. The canonical reference for the AI story.
Read →Standalone npm package: @quilt/ai. 4 provider implementations, AIEngine with caching + cost tracking, 6 examples.
Generate code with a model cell. Add a runner cell. The cells form a complete code generation pipeline. Formulas can be model outputs.
Try it →# A Quilt cell that uses z.ai GLM 5.3 - id: ai.glm kind: ai.llm provider: zai model: glm-4.5 prompt: "Translate to French: {{input.text}}" # A Quilt cell that uses Cloudflare Workers AI (free, runs on edge) - id: ai.embed kind: ai.embed provider: cloudflare model: "@cf/baai/bge-base-en-v1.5" input: ai.glm # A Quilt cell that uses a Kimi model - id: ai.kimi kind: ai.llm provider: kimi model: moonshot-v1-8k prompt: "Summarize: {{input.text}}" # And one that uses DeepSeek (cheap, fast, wide view) - id: ai.deepseek kind: ai.llm provider: deepseek model: deepseek-chat prompt: "Generate 3 ideas for: {{input.text}}"
The 12-month plan is below. The 5-year vision is the cell model everywhere: microcontroller, browser, OS, silicon.
Open Quilt Live. Click "Add cell". Type a value. Add a formula. Watch it compute. Save your work. Download the file. Open it offline. It's all yours.