NovaFabric

v0.58.0 — self-hosted, open source

The time machine for AI systems.

Capture every AI run. Replay any of them. Diff what changed. Trace what depended on what. Export signed evidence. Self-hosted. Open source. CLI-first.

pip install novafabric
See an example run

This site is documentation. The CLI is the canonical interface — a read-only browser viewer is planned for v1.x.

A 90-second story

Yesterday it worked. Today it doesn't. Now what?

01 — yesterday

It worked.

Your code-review agent shipped a thoughtful review on a 200-line PR. Captured automatically — every model call, every tool, every env fact.

nova capture python review.py
02 — today

It fails.

Same PR, different review — shallower, missing the bug yesterday's run flagged. Something upstream changed. The prompt? The model? A tool version?

nova diff RUN_A RUN_B
03 — answer

Replay yesterday's run.

Re-execute it under the new prompt. Lineage shows the prompt changed from @0.1.0 to@0.2.0. Eval suite caught it. Promotion blocked. Production stays on v0.1.0.

nova replay RUN_A --mode mocked

The five primitives

Five primitives that compose into a time machine.

Each one stands alone. Together they capture, replay, and prove what your AI system did — yesterday, last quarter, two years ago.

01

Asset Registry

Identity layer for every model, agent, prompt, tool, dataset. Eval-gated promotion from development → production.

Explore
02

Run Capsule

A self-contained record of one AI run. Inputs, outputs, model calls, tools, environment — schema-validated.

Explore
03

Replay

Reconstruct yesterday's run today. Four modes — exact, mocked, semantic, forensic — with honest guarantees.

Explore
04

Lineage

Causation graph linking runs, assets, and artifacts. Provenance, blast radius, replay chain — three queries.

Explore
05

Evidence Bundle

Signed, portable proof of what happened. in-toto DSSE, ed25519 — verifiable in your browser, right now.

Explore

Honest limitations

What we don't promise.

Not byte-exact replay of remote LLMs.If you call OpenAI or Anthropic, NovaFabric replays your run from captured records, not from re-issuing the API call. Four replay modes (exact,mocked,semantic,forensic) make the guarantee explicit per run.

Not a hosted SaaS.v0.x runs entirely on your machine. SQLite for the registry and lineage. Local files for capsules. v0.7 introduces an optional server mode; v1.x adds a local read-only browser. Until then, this site is the only browser experience.

Not an evaluation harness.NovaFabric records what happened so your evals can answer "did the system regress?" Eval suites are a separate primitive (see the registry demo); we don't ship benchmarks.

Try it on your own code.

Three commands. No accounts. No telemetry. Capsules live in.novafabric/runs/ on your machine.

pip install novafabric
nova capture python your_agent.py
nova replay <run-id> --mode mocked
Install instructions·GitHub·Concepts in depth