Test an MCP server before connecting it to an AI agent.

MCP Server Auditor

• ✓ Discover exposed tools and capabilities
• ✓ Identify risky, mutating, or destructive operations
• ✓ Detect protocol, schema, and reliability issues
• ✓ Receive deployment-readiness findings
• ✓ Get structured machine-readable evidence
• ✓ Batch-audit up to 25 MCP servers

MCP Server Auditor evaluates public, unauthenticated Streamable HTTP endpoints through Sentinel Verify. Each target is checked locally against API-IFY's public-network policy before it is sent to Verify.

Quick start

The Sentinel-owned endpoint below is the controlled, discovery-only qualification target used by the Store prefill.

{
  "servers": [{"url": "https://mcp.sentinelsignal.io/audit-fixture/mcp"}],
  "depth": "standard",
  "includeEvidence": false
}

The Actor performs MCP initialization and discovery, evaluates the advertised surface, and returns one structured item per unique target. A successful result has this shape:

{
  "status": "success",
  "result": {
    "capabilities": {"toolCount": 3, "resourceCount": 0, "promptCount": 0},
    "findingCount": 0,
    "summary": {"score": 65.1, "readiness": {"label": "Safe for evaluation"}}
  }
}

See examples/sample-input.json and examples/sample-output.json for the complete contracts. The checked-in output was produced through the Actor's real service path against the controlled endpoint and Verify 1.0.703. Scores and findings depend on the endpoint's current behavior and Verify rules.

Input and batching

servers accepts one to 25 objects with a public HTTPS url. Duplicate URLs are collapsed and produce one result with a warning. depth is currently fixed to standard. Set includeEvidence to false for a smaller result.

Output and failure behavior

Results use the stable API-IFY schemaVersion: "1.0" envelope. The default dataset contains audit or failure records, while RUN_SUMMARY in the default key-value store records processing, delivery, billing, dependency, and budget totals.

restricted means API-IFY rejected the target locally. unreachable means the public endpoint could not complete MCP discovery. unsupported covers unsupported target behavior. dependency_unavailable means Verify was temporarily unavailable. failed is reserved for unexpected audit failures. Failures are isolated per target.

Pricing

One mcp-audit event is charged only when a successful or useful partial audit is accepted into the dataset. Restricted, invalid, unreachable, unsupported, dependency-failure, and failed outcomes are uncharged. Work does not start when the caller's remaining event budget cannot cover another audit. The current event price is displayed by Apify.

Limitations

• Public, unauthenticated MCP endpoints only.
• HTTPS on port 443 only; credentials and fragments in URLs are rejected.
• Streamable HTTP transport only.
• Private, loopback, link-local, metadata, reserved, and mixed public/private DNS targets are rejected.
• Maximum 25 submitted targets and four concurrent audits per run.
• Audit availability depends on Sentinel Verify and the target MCP server.

Privacy and security

The target endpoint is sent to Sentinel Verify for analysis after local public-network validation. Results are stored in the customer's Apify run dataset and summary store. Do not place credentials in target URLs or query parameters. Published datasets, key-value records, and structured logs redact authorization values, secrets, private keys, and secret-like query parameters.

VERIFY_API_KEY is an operator-managed Actor secret. VERIFY_BASE_URL and VERIFY_TIMEOUT_SECONDS are operator settings, not Actor inputs.

Support

For a reproducible support request, provide the Apify run ID, result itemId, and status. Never send credentials or unredacted URLs through support. Contact Sentinel Signal Systems through the support link on the Actor page.