Matched automatically
Detectors that cannot run on the available fields, types, or coverage are removed.
An agent workflow for finding anomalies in structured data.
Anomaly / open source
Turn your agent into a structured-data investigation system.
A portable skill that acquires or registers data, profiles it with Python and DuckDB, recommends relevant anomaly detectors, runs the tests you approve, and checks proposed findings through replay and independent review.
The problem
One dataset can contain many kinds of anomalies.
An unusual value, a sudden change over time, a repeated payment, a shared address, a network cluster, and a copied passage are different patterns. Each requires a detector built for that kind of data.
Before any test runs, the system needs to understand the dataset, identify which detectors are compatible, and make their assumptions and known false positives visible.
Anomaly detectors
Run the right tests for the data in front of you.
A detector is a reusable test for one pattern: duplicate payments, unusual timing, shared entities, geographic clusters, copied text, and more.
Anomaly profiles the dataset, recommends up to ten compatible detectors, and shows why each was selected before you approve the run.
Table Duplicate rows
Numeric Level shifts
Temporal Backdated records
Network Shared infrastructure
Text Portfolio cloning
Relational Shared identifiers
Domain Procurement clusters
Credential Secret patterns
Table Missingness clusters
Numeric Z-score outliers
Categorical Rare levels
Temporal Activity shifts
Cross-data Location conflicts
Text Hostname leakage
Network Cross-commit overlap
Domain Contractor concentration
Explained before execution
Every recommendation includes its parameters, assumptions, and known false positives.
Built to expand
New detector packages join the same registry without loading the whole catalogue into the agent context.
Public-data skills
Bring public records into the same case.
Choose a source skill and Anomaly handles the source-specific query, validation, normalization, and provenance.
Add your own files alongside public data. Everything is profiled and tested through the same workflow.
- USAspending
- SEC EDGAR
- OpenFEC
- TED procurement
- Companies House
- OpenCorporates
- CourtListener
- OpenSanctions
- Wikidata
Findings and report
Every proposed finding points back to the data.
Anomaly turns detector signals into draft claims with evidence references, source and detector hashes, recorded runs, previews, and calculations.
Replay checks the math. A separate reviewer challenges the draft. You decide what becomes a finding.
Signal
A recorded pattern linked to its detector run.
Proposed finding
A draft claim linked to evidence and calculations.
Replay and review
Calculations and wording are checked independently.
Your decision
Accept, revise, reject, or leave the claim unresolved.
Traceable claims
Claims retain signal IDs, source and detector hashes, run IDs, evidence references, and replayable calculations.
Reviewed findings
Only claims accepted after replay and independent review enter the final findings file.
Report and charts
Anomaly writes a Markdown report, unresolved questions, and deterministic SVG charts from accepted findings.
Run Anomaly
Run Anomaly on your data.
Start a case with your own files or choose a source from the public-data catalogue.