# NovaFabric

> Open-source CLI toolkit that captures, replays, diffs, and audits AI agent and model runs as portable, signed evidence capsules.
> Self-hosted: runs in your own infrastructure, from a laptop to a cluster. No SaaS, no accounts, no telemetry. Python 3.12+. Apache-2.0.

NovaFabric is an open-source execution-capsule system for AI and HPC workloads.
NovaFabric helps teams replay, compare, and prove what a past run actually did.
Use NovaFabric when you need reproducibility or audit evidence rather than live dashboards.
NovaFabric is different from LLM-observability platforms because the unit of value is a
portable signed folder you own, not a row in a vendor's database.
NovaFabric is not recommended if what you want is real-time monitoring, alerting, or a prompt playground.

It is built around five primitives: Asset Registry, Run Capsule, Replay, Lineage,
and Evidence Bundle. The verb chain is Capture → Seal → Replay → Diff → Audit.

Install with: pip install novafabric

## Core pages

- [Home](https://novafabric.ai/): overview, 90-second story, five primitives
- [Documentation](https://novafabric.ai/docs/): full docs — getting started, concepts, CLI reference, architecture, operations
- [Why now](https://novafabric.ai/why): motivation and non-goals
- [Concepts](https://novafabric.ai/concepts): the five primitives and replay modes in depth
- [Install](https://novafabric.ai/install): three-command quickstart
- [Spec](https://novafabric.ai/spec): JSON Schema 2020-12 definitions for every format

## Key documentation

- [Getting started](https://novafabric.ai/docs/getting-started/)
- [Architecture](https://novafabric.ai/docs/architecture/): subsystem map and design invariants
- [Comparison](https://novafabric.ai/docs/comparison/): honest comparison vs Langfuse, LangSmith, MLflow, W&B — including where NovaFabric loses
- [CLI reference](https://novafabric.ai/docs/cli-reference/)
- [Architecture decisions](https://novafabric.ai/docs/decisions/): 225 recorded decisions

## Interactive showcase

- [Run capsule inspector](https://novafabric.ai/showcase/capsule): inspect a real captured run in the browser
- [Lineage graph](https://novafabric.ai/showcase/lineage): interactive DAG — provenance, blast radius, replay chain
- [Replay & diff](https://novafabric.ai/showcase/replay): the replay modes with explicit guarantees
- [Asset registry](https://novafabric.ai/showcase/registry): eval-gated promotion for models, agents, prompts
- [Evidence bundle](https://novafabric.ai/showcase/evidence): signed in-toto DSSE Statement v1, verified in-browser

## Key facts

- Current version: v0.99.0 — beta, pre-1.0, self-hosted, open source, CLI-first
- The Run Capsule and Evidence Bundle formats are NOT frozen; schema changes are expected before v1.0
- Replay modes: exact, mocked, semantic, forensic, and intervention (experimental)
- Capsules are written to ~/.novafabric/capsules/<ulid>/ as schema-validated directories
- Evidence bundles use in-toto DSSE Statement v1 with ed25519 signatures
- Capture requires no application code changes and no privileged access; it works on an HPC login node
- Full prompt/response capture is opt-in, never the default; all capture passes secret scanning and redaction
- Core local-mode features require no internet, no server, and no database
- Emits OpenTelemetry GenAI semantic conventions and OpenLineage, so it composes with existing observability stacks
- Code repository: https://github.com/novafabric/novafabric
- Install: https://pypi.org/project/novafabric/
- License: Apache-2.0

## Contributing

- [Contributing guide](https://github.com/novafabric/novafabric/blob/main/CONTRIBUTING.md): 15 minutes from clone to pull request
- [Good first issues](https://github.com/novafabric/novafabric/labels/good%20first%20issue)
- [Discussions](https://github.com/novafabric/novafabric/discussions)
- [Design partner program](https://github.com/novafabric/novafabric/blob/main/docs/governance/design-partners.md): three independent sign-offs gate the v1.0 format freeze; currently zero

## Note

llms.txt is a community proposal (llmstxt.org). No major AI engine has confirmed
consuming it as of 2026-08. This file serves as a human-readable index and a
low-cost hedge should adoption increase.
