WaveGuardClient is a Python software development kit for the WaveGuard anomaly detection API. WaveGuard is a general-purpose anomaly detection API whose core feature is that it is based on wave physics rather than machine learning. Users can send it arbitrary data—server metrics, financial transactions, log files, sensor readings, time series, etc.—and get anomaly scores, confidence levels, and explanations of which features triggered the alert. It requires no training pipeline, no model management, and is stateless; just one API call.
In terms of key capabilities, WaveGuard uses GPU-accelerated wave physics simulation under the hood. Data is encoded onto a 64³ lattice, and the simulation runs on the GPU via coupled wave equations. Normal data produces stable wave patterns, while anomalous data produces divergent patterns. A 52-dimensional statistical fingerprint comparison is performed between training data and test data. Everything is torn down after each call, and nothing is stored. Advantages over machine learning include: no training data requirement (2 or more samples are enough), no model drift, no retraining, and no hyperparameter tuning; the same API call works for structured data, text, numbers, and time series. Version v2.2 adds multi-resolution scoring. In addition to the global fingerprint distance, it also tracks the local lattice energy of each feature, making it possible to capture subtle per-feature anomalies that the global averaging in v2.1 missed (for example, 3 out of 10 IoT sensors drifting), improving IoT F1 from 0.30 to 0.87.
In terms of usage, the SDK's only dependency is the requests library, and all physics computations run on the server-side GPU. Users first need to obtain a free API key on RapidAPI. The core call is wg.scan(training, test, encoder_type=None, sensitivity=None), which returns a ScanResult containing .results (per-sample results) and .summary (summary results). It also provides wg.health() and wg.tier() for health checks (no authentication required) and retrieving subscription tier information. v3.3.0 introduces advanced intelligence methods, including wg.counterfactual(...), wg.trajectory_scan(...), wg.instability(...), wg.phase_coherence(...), wg.interaction_matrix(...), wg.cascade_risk(...), wg.mechanism_probe(...), wg.action_surface(...), wg.multi_horizon_outlook(...). These methods map directly to the /v1/* intelligence endpoints and return raw JSON payloads for maximum compatibility with rapidly evolving server-side response schemas. Data shapes are auto-detected and require no configuration.
In terms of use cases, the official documentation provides multiple industry examples: detecting compromised servers, flagging fraudulent transactions, and catching security incidents in logs, all using the same client and scan() call. There is also a crypto risk scanner, CryptoGuard, powered by WaveGuard, which was backtested against 7 historical crashes (LUNA, FTX, Celsius, 3AC, UST, SOL/FTX, TITAN). WaveGuard issued a CAUTION warning for FTT (the FTX token) on October 16, 2022, 23 days before 94% of the crash, while z-score analysis showed nothing abnormal. At the same recall rate, it had 5 times fewer false positives than the statistical baseline. The SDK can also be used as an MCP server for AI agents such as Claude Desktop, providing MCP tools such as waveguard_scan. In addition, WaveGuard is considered an alternative to Azure Anomaly Detector (retiring in October 2026).
Limitations and notes: The official materials present benchmark tests comparing v2.2 with scikit-learn in 6 real-world scenarios; WaveGuard won 4 out of 6 scenarios, with an average F1 of 0.76; it ranked first in F1 score on 12 public benchmark datasets (see HuggingFace for the full comparison). The project uses the MIT license. Detailed information such as specific pricing, costs, and platform support scope is not yet fully provided in the official materials and should be referenced on the RapidAPI page.
