# KryptosAI/mcp-observatory [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/KryptosAI/mcp-observatory  
**GitHub Stars:** 135  
**npm Downloads (last month):** 2247  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/kryptosai-mcp-observatory

## Description
Regression testing for MCP servers. Auto-discovers servers from Claude configs, checks capabilities, invokes tools, detects schema drift between versions, and recommends new servers based on your environment. Works as both a CLI and an MCP server.

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

```json
"mcpServers": {
  "mcp-observatory": {
    "command": "npx",
    "args": ["-y","@kryptosai/mcp-observatory@latest"]
  }
}
```

## Documentation

## What KryptosAI/mcp-observatory does

KryptosAI/mcp-observatory evaluates MCP servers as test targets rather than using them as application dependencies. It can inspect configured servers, connect to a selected command, enumerate capabilities, invoke tools, and report failures or security-related findings. Regression workflows cover compatibility, tool behavior, schema changes, record/replay/verify checks, and response timing.

Results include a target, grade, findings, and a saved receipt. Receipts package evidence, a verdict, an action, and reproduction commands. The project also supports risk graphs, health score badges, GitHub pull request comments, and normalized SARIF output for GitHub Code Scanning. An MCP server mode lets agents inspect other MCP servers directly.

## How it works

Run the CLI against a configured server or pass a startup command explicitly. With no configured target, the first scan uses an included example. The scanner stores local history and can generate HTML reports and JSON downloads through its dashboard. `audit` supports profile-mapped checks such as `nsa-mcp`, while `setup-ci --all` can generate a complete CI workflow.

The project can also apply defensive controls around discovered servers. `protect` rewrites client configurations for a default-deny setup and supports undoing those changes; HTTP and SSE servers are left untouched. `enforce --start-proxy` creates a Seatbelt policy from scan results and starts the proxy. These controls are separate from the local scanning workflow.

## Setup and configuration

Node.js 20 or newer and npm are supported on macOS, Windows, and Linux. The quickest installation runs the package directly:

```bash
npx -y @kryptosai/mcp-observatory@latest
```

Other supported installation paths include Homebrew, Docker, and the repository's shell installer. Docker scans of host-local servers require network access and, when local files are involved, an appropriate volume mount and working directory.

To use the server through an MCP client, start its stdio mode with:

```bash
npx -y @kryptosai/mcp-observatory@latest serve
```

CI can use the repository action or a generated workflow. A direct test looks like this:

```bash
npx -y @kryptosai/mcp-observatory@latest test npx -y @modelcontextprotocol/server-everything
```

Telemetry is disabled by default. `telemetry preview` shows the fields, and `telemetry enable` opts in. Local scans, history, evidence export, and CI operation do not require a ClearFrame account; the optional `clearframe` command prepares a private instruction file and a read-only CLI command for review.

## Tools and capabilities

- Discover MCP servers from Claude and other supported client configuration locations.
- Check server capabilities and invoke exposed tools.
- Detect schema drift and run compatibility or regression checks.
- Generate MCP receipts, reports, JSON exports, SARIF, and CI findings.
- Record, replay, and verify MCP interactions.
- Audit against a named profile, including `nsa-mcp`.
- Review local scan history through a dashboard.
- Produce risk graphs and action outcomes such as allow, gate, rerun, quarantine, or escalate.
- Learn from explicitly integrated usage data, including sanitized invocation evidence, intent grouping, dissatisfaction flags, and regression-case export.

## Limitations and notes

The `clearframe` workflow is optional and does not launch ClearFrame or upload evidence automatically. ClearFrame itself is currently available for Apple Silicon Mac, with separate account and usage terms. Usage learning requires an explicit local integration; observed execution and inferred breakdown signals are kept separate, and the review screen runs in a loopback-only browser context.

`protect` does not modify HTTP or SSE servers. The Docker workflow needs explicit access to the host network and local project files when scanning host-local targets. The project is free, open source, and MIT licensed.

_Full upstream README: https://allmcps.com/mcp/kryptosai-mcp-observatory/readme_