Where ideas come alive.
One day the thought just hit me: why not build an app builder that actually works, with every feature I'd want from one? You describe an application; Vivarium questions you until the idea is specified well enough to build, shows you a preview to approve, then generates a working full-stack app phase by phase, on your own machine.
Most AI builders fail in the same two places: they stop asking questions too early, and the model that writes the code also writes the tests that check it. Vivarium is designed around avoiding both.
From a paragraph to a running app.
The three rules it's built around.
A classifier picks the kind of app (dashboard, marketplace, chat, workflow…), which selects a checklist. The interview ends when the checklist is satisfied, not when the model feels done. A model that asks five questions and declares itself satisfied builds the wrong thing.
A planner writes the tests first. Builder agents write the app and are blocked from the tests folder by filesystem permissions and pre-commit checks. If one agent writes both, it writes tests that pass.
In development builds, every element is stamped with the file and line that made it. Right-click something in the running app and the model gets the real source location, not a guess.
Underneath all three is an append-only project memory. Every decision you make is recorded with where it came from (you, a default, an inference, an agent), and your own choices are binding: no agent may overwrite them. Secret values never enter a model's context.
The parts with no AI in them come first.
Four subsystems carry most of the system's correctness guarantees, and none of them use a model: the store (project memory as append-only JSONL in git, with SQLite rebuilt from it), the templates (hand-built reference code every project starts from), the index (how the code fits together), and the server (a read-only view other AI tools can query).
“The server reads. The supervisor executes. The store records. The template constrains.”
Phase 2 of 8, underway.
- phase 0 · doneRecord store: append-only memory with provenance
- phase 1 · doneTemplates and the local dev runtime
- phase 2 · in progressIndexer: structural and semantic code maps
- phase 3Read-only MCP server
- phase 4Three-view preview generator
- phase 5Planner agent: contracts and tests
- phase 6Builder agents: run, verify, repair
- phase 7Build presentation: live event stream