Aurora is a local-first, open-source, verifiable quantitative analysis tool, positioned as "glass-box quantitative intelligence." Its core philosophy is: cloud-based large language models tend to guess, whereas Aurora reaches conclusions through computation and verification; every relationship is calculated rather than predicted, and every conclusion is annotated with its source.
Purpose Aurora serves two types of users: first, analysts, quantitative researchers, scientists, and engineers who need to analyze hard data; second, developers of AI systems and agents who cannot tolerate hallucinations. It is used for quantitative analysis tasks such as anomaly detection, causal inference, forecasting, physical law discovery, and data quality pre-checks.
Main capabilities
- Six analytical perspectives: overview, anomaly, state, pattern, forecast, physics.
- More than 24 research-grade methods: Isolation Forest, robust z-score (Hampel), Granger, HMM Baum-Welch, persistent homology, SINDy, Gaussian processes, mutual information, VAR, DTW, BOCPD, Robust PCA, EMD, Kalman, spectral entropy, etc.
- Knowledge-base-supported synthesis: every sentence of "What This Means" cites seed:* entries from a publicly licensed knowledge base.
- Pre-check data quality: schema validation, missing-pattern detection, irregular sampling checks.
- Causal inference (do-calculus): Pearl-style backdoor identification and adjusted OLS estimation, supporting counterfactual queries.
- Streaming/continuous mode: point to a directory, refresh conclusions when new data arrives, and deduplicate by conclusion.
- Composable conclusions: the results of a previous run can serve as the prior for the next.
- Multi-dataset connection: compare shared keys, schema compatibility, and inheritance candidates between two runs.
- Run library: pin runs across sessions, A/B comparison, and share as .aurora.json packages.
- Custom knowledge base ingestion: supports PDF/TXT/MD folders.
- Knowledge base package marketplace: install curated domain knowledge packages (climate, finance, biomedical, industrial).
- Package provenance: triple check of integrity, Ed25519 signature, and trusted signer registry.
- Calibration engine: provides empirical false positive rates on signal-free data for the change point detection family (BOCPD, CUSUM, PELT).
- Honest disclosure: sampling, timeouts, and skipped methods are clearly displayed, and it will not pretend to execute.
How to use Desktop application: download the installer for the corresponding operating system and double-click to run. On first launch, it will guide you through a small knowledge base seed, and thereafter runs fully offline. Supports drag-and-drop of CSV, Parquet, JSON, and XLSX files. Running from source: run studio_api.py, open http://127.0.0.1:8000 , you can click the demo or drag in data. The first run will download about 50 MB of knowledge base seeds. Docker: run with one command, with state persisted in the host's ./aurora-data/ directory. MCP and SDK: Aurora Cortex provides an MCP server and Python SDK, which can be integrated with Claude Desktop and others; examples are located in examples/mcp-claude-desktop/. Decision contracts: you can configure automation such as "trigger when more than 3 severe anomalies occur"; examples are located in examples/decision-contracts/, supporting linked demos such as Discord, Slack, OBS overlay cards, and smart plugs.
Applicable scenarios
- Anomaly and predictive analysis of hard data such as ambient air quality and industrial bearings
- State identification and causality testing of financial and economic data
- Physical law discovery and pattern recognition in scientific research data
- AI agents calling Aurora for verification before generating quantitative conclusions
- Compliance scenarios requiring local deployment and data not leaving the machine
Limitations The desktop version is currently not code-signed, so Windows SmartScreen and macOS Gatekeeper will prompt "unknown publisher"; this is expected behavior. There are no Intel-Mac (x86_64) or Linux AppImage builds; Intel-Mac users can run from source. The backend needs 5–10 seconds of warm-up on first startup; analysis of large datasets takes a long time and results are returned all at once; duplicate files complete instantly due to caching. The knowledge base is small on a fresh installation and grows over time. Some connectors (Kafka, Postgres CDC) have dependencies that you must satisfy yourself. Pricing, paid plans, specific performance metrics, and more platform support information are not yet provided.
