The first one
Idea Garden
The first form and board, and the walkthrough for newcomers: an empty file, editing by hand, a first shape built by an agent, a CSV import, focused screens, and recovery offline.
Concept and philosophy
Nendo tests one idea: that an application can be changed in place instead of regenerated. This page sets out the idea, the rules that protect it, and what would prove it wrong.
AI can write a small application in minutes, and each one becomes another silo. The data is tied to throwaway code, every change means generating the code again, and you can't see how the schema changed, what happened before, or how to go back. When the code stops working, the data is stuck with it.
Low-code platforms solve part of this but bring a lock-in of their own: your application lives on their servers, in their format. You can't keep it as a file to copy, inspect and open offline ten years from now.
Treat an application like clay. A single SQLite file, the .nendofile, holds the schema, the data, the screens and the full history, with a stable ID for everything in it. People and coding agents change it in place using typed operations. Nothing is generated and nothing is built.
Nendo (粘土) is Japanese for clay, and the name is the idea: clay holds a shape and can take a new one without being thrown away.
The axioms
Quoted from the vision document. Breaking one takes a formally accepted architecture decision; a good argument is not enough.
A .nendo file with no user schema is a working application, not an error state.
No application content and no agent can remove the route back to the data.
The table works before the form exists, and after it breaks.
Native UI, web Workbench and MCP adapter call the same typed application services.
No SQL, no SQLite handle, no database path, no generic host invocation.
Validation happens on a clone with no authority over the active file.
Every operation declares its reversibility class. There is no universal undo.
How it could fail
The idea holds if an agent, using nothing but the MCP interface, can build an application nobody designed Nendo around, and a person can read that change and understand it before accepting it.
It fails if the built-in table is all anyone needs. If custom screens add nothing that Studio doesn't already give you, the project should change course or stop.
The evidence so far
One application shows the approach can work. Four different ones begin to show it wasn't tailored to the first. The last two were built from an empty file by an agent working only through MCP.
The first one
The first form and board, and the walkthrough for newcomers: an empty file, editing by hand, a first shape built by an agent, a CSV import, focused screens, and recovery offline.
Nothing special-cased
Built to check that nothing in the path was specific to Idea Garden. The same building blocks make a different application.
Built over MCP alone
Record pages, related lists, filters, summary tiles and multi-step commands, all built by an agent from an empty file. It learned what it could build from the MCP interface, not from the source code.
The demonstration
An operations room for a fictional habitat: a front page, record pages with tabs, a schematic drawn by a custom view, commands, an automatic trigger, and an agent’s change reviewed live.
Scope
The out list isn't a to-do list. Each item was left out on purpose, often to keep what is in small enough to reason about.

A new user creates an empty file, shapes or imports a modest dataset, uses it productively through Studio, has an external agent build a materially better focused surface, reviews and accepts the semantic change, restarts offline, and recovers the data without that surface.