The experiment lab
for the agentic era
NorviaLabs ships a family of focused open-source experiments for AI builders — runtimes, observability, research harnesses, and developer tooling. Small sharp ideas. Public by default. MIT throughout.
What is NorviaLabs
An open-source lab for experiments that should have been public sooner
AI teams ship agents faster than the tooling underneath them can keep up. Evaluations are fragile, observability is afterthought glue, and half the “research” never leaves a private notebook.
NorviaLabs publishes that missing layer as focused, MIT-licensed experiments. Sometimes it is a runtime primitive — a sandbox, a memory shard, a trace kit. Sometimes it is a research harness or a small tool that makes the next experiment cheaper to run. Each one does one thing well and lives in the open.
# Small imports. Sharp ideas. Ship in the open. from tracekit import Tracer from sandbox_lite import Sandbox from costmap import CostMap tracer = Tracer() with Sandbox(timeout=30) as box: result = box.run(agent_step) tracer.span("agent_step", result) CostMap.from_run(tracer).print() # where did tokens go?
The portfolio
Open experiments, one thesis
Each experiment encodes one sharp idea, ships its own docs, and is free to fork. Filter by what you are building.
No experiments in this family yet.
See all experiments →What we build
Five families across the experiment stack
Every experiment fits into one of them.
Find a tool
If you need this, reach for that
A decision table across the portfolio. Match your problem to the experiment built for it.
| If you need to… | Reach for | Family | Stack |
|---|
How we build
The same principles run through every experiment
Public by default
Experiments start open. Source, issues, and notes ship together so others can reproduce, fork, or kill the idea honestly.
One sharp idea each
Small products that do one thing well and compose. No kitchen-sink frameworks — just a clear README and a hard boundary.
Measure what matters
Benchmarks, traces, and cost maps are first-class. If we cannot measure it, it is not ready to recommend.
MIT, no lock-in
Every experiment is MIT licensed. No accounts, no telemetry tax, no closed tiers — the library or CLI is the whole thing.
Start with a single experiment
All NorviaLabs projects are open source and MIT licensed. Every README, changelog, and benchmark is in the open.
FAQ