Strata.
A modest local model can do real work if the system around it supplies the rigor it lacks. Strata builds that system and measures whether it works.
- Status:
- prototype (Working prototype, paused)
- Built with:
- Python, FastAPI, React
- Licence:
- Apache-2.0
What it is
Strata is a local-first agent harness. A FastAPI backend and a background worker run tasks through decomposition, research, implementation and judging, and record every step as an attempt with an outcome, a reason and a resolution.
It has two model tiers on purpose: a local agent that helps the user, and a cloud trainer that improves the agent. A React workbench shows chat, live lanes, tasks, attempt history and telemetry.
The thesis
The usual move is to use a bigger model. Strata moves the intelligence into the system around the model (rigor, memory, decomposition and verification), then checks whether that worked, and whether the small model has started improving itself.
What works today
- It runs end to end on a laptop, and its eval-driven self-improvement loop exists and is instrumented.
- A verifier that checks every output while a lane has little evidence, then relaxes as clean evidence accumulates.
- Context loads are costed in estimated tokens, and each model endpoint paces its own requests.
- A workbench demo mode that replays a recorded sample run with no backend or model, labelled as demo data.
- The local API checks the Host and Origin of every request and needs a per-start token for writes.
Limitations and scope
- On small coding tasks graded by their unit tests (October 19, 2026, a 4B local model on a CPU-only host), no improvement was measured. Strata as shipped wrote no code, and with a reviewed fix it still produced no passing code.
- Self-improvement has not been measured: the promotion cycle scores proposals with a cloud judge model and cannot run with only a local model.
- A working prototype, paused after a 12-day build in March and April 2026. Expect rough edges, an evolving schema and an operator workbench rather than a product.
- At the published snapshot, 15 of 315 tests fail; they predate the snapshot and were left as found.
- No user accounts. It is meant to run on your own machine for you alone, and does not protect against other programs running as you.
Source
The code is public. Everything on this page comes from Strata's own README; the README is the place to check what has changed since.