MCP server security: redacting sensitive data

Tool calls often carry sensitive data: emails, tokens, query contents. You can still measure usage without shipping any of it, as long as redaction happens locally, before anything leaves your process.

Redact in-process

Good MCP analytics strips sensitive fields on your server, not on a remote pipeline. You name the paths to remove, and the raw values never travel.

withTrackMCP(server, {
  apiKey: process.env.TRACKMCP_KEY,
  redact: ["args.email", "args.api_key", "result.raw_response"],
});

Prefer shape over content

You rarely need the actual values to learn from usage. Knowing that a call happened, which tool, which client, and whether it succeeded is enough for adoption, workflows, and reliability.

Other safeguards

• Sample high-volume calls to reduce data footprint
• Fail open: capture must never block a tool call
• Scope API keys to a workspace and keep them server-side

Done right, you get the full analytics picture while sensitive data stays inside your infrastructure.