# gpartin/WaveGuardClient [Health: Active]

**Category:** 🧮 Data Science Tools  
**Repository:** https://github.com/gpartin/WaveGuardClient  
**GitHub Stars:** 4  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/gpartin-waveguardclient

## Description
Physics-based anomaly detection via MCP. Uses Klein-Gordon wave equations on GPU to detect anomalies with high precision (avg 0.90). 9 tools: scan, fingerprint, compare, token risk, wallet profiling, volume check, price manipulation detection.

## Tools
Capabilities this server exposes over MCP:

- **waveguard_scan** — Detect anomalies in any structured data
- **waveguard_scan_timeseries** — Auto-window time-series and detect anomalous segments
- **waveguard_health** — Check API status and GPU availability

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "waveguardclient": {
    "command": "uvx",
    "args": ["--from","WaveGuardClient","waveguard-mcp"]
  }
}
```

## Documentation

## What gpartin/WaveGuardClient MCP server does

The gpartin/WaveGuardClient MCP server gives an MCP-compatible agent access to WaveGuard, a general-purpose anomaly detection API. It accepts structured data and time-series data, then returns anomaly scores and related detection results. The README also describes support for numeric and text data through the underlying WaveGuard client.

The detection process is stateless. Data is sent to the API for processing, and the README states that nothing is stored after each call. The compute runs server-side on GPU infrastructure, so the client does not need CUDA or a local model runtime.

## How it works

The gpartin/WaveGuardClient MCP server passes agent requests to the WaveGuard API. According to the project documentation, the API encodes data onto a 64³ lattice and runs coupled wave-equation simulations on a GPU. Normal inputs produce stable patterns, while anomalous inputs produce divergent patterns. The service also compares a 52-dimensional statistical fingerprint between training and test data.

For ordinary scans, callers provide examples representing normal or training data and test records to evaluate. The API can return anomaly scores, confidence values, and explanations of which features contributed to an alert. Time-series requests can be divided into windows automatically so the service can identify anomalous segments rather than treating an entire series as one record.

## Setup and configuration

The Python package is installed with:

```bash
pip install WaveGuardClient
```

The README shows the client being initialized with an API key obtained through RapidAPI. The supplied material does not document a specific environment-variable name or a standalone MCP launch command, so configuration should follow the repository's current MCP or client instructions rather than assuming one.

Because processing is performed by a remote API, network access and a valid API key are required. The only Python dependency identified in the README is `requests` for the SDK. Local GPU hardware is not required for the documented client flow.

## Tools and capabilities

The documented MCP tools are:

- `waveguard_scan`: Detects anomalies in structured data.
- `waveguard_scan_timeseries`: Automatically windows time-series input and detects anomalous segments.
- `waveguard_health`: Checks API status and GPU availability.

The broader API examples cover server metrics, financial transactions, log records, sensor readings, and other structured or text inputs. The project describes the same underlying scan operation as usable across these data types, provided the caller can define representative normal data.

## Limitations and notes

The gpartin/WaveGuardClient MCP server depends on the remote WaveGuard API rather than performing detection locally. API availability and GPU status can be checked with `waveguard_health`. An API key is required, and the README points to RapidAPI for obtaining one.

The published benchmark material reports different results by scenario, including weaker performance for the listed time-series benchmark than for several other categories. These figures are benchmark results, not a guarantee for a particular dataset. The project also describes a minimum of two samples as sufficient for its approach, but callers still need suitable examples of normal behavior for meaningful anomaly detection.

The README identifies Claude Desktop as an MCP target. It does not provide enough information here to confirm compatibility with other clients or to specify the exact MCP startup command. The gpartin/WaveGuardClient MCP server should therefore be evaluated with the repository's current configuration instructions before deployment.

_Full upstream README: https://allmcps.com/mcp/gpartin-waveguardclient/readme_