The research goal: can a piece of work be split across many AI models, local and cloud, without the context dying every time the work moves between them? Losing context on a model swap is the failure that makes most multi model setups unusable, and this project exists to solve it.
It is internal infrastructure. It is not a product and it is not sold separately. It exists because building the rest of the lab's projects at this pace is not possible by hand, and everything learned here feeds the production suite.
Work is dispatched across three role slots rather than thrown at one general model. An Architect plans, reviews, and verifies meaning. An Implementer makes the bulk code changes and runs the tools. An Adaptive role scans, lints, and classifies, and is deliberately read only so it can never make broad changes on its own. Each role maps to the class of model that is good at it, which is also what keeps the cost down.
A shared context store keeps the thread intact when a task moves between models. A code knowledge graph built on Neo4j and Tree-sitter gives the roles structural understanding of a repository instead of whatever fits in a prompt window. Local models run through Ollama, so the free tier costs nothing to operate.
It runs as a local first system: a Python backend, a React frontend, and a native desktop shell, all on the user's own machine.
What the lab built
At a glance
A piece of work arrives with its context
Plans, reviews, verifies meaning
Makes the changes, runs the tools
Scans and classifies, read only
The thread survives every model swap
