Skip to content

quickstart.md

Latest commit

 

History

History
35 lines (27 loc) · 1.66 KB

quickstart.md

File metadata and controls

35 lines (27 loc) · 1.66 KB
Raw

Quickstart: Output Sanitisation (Track B)

Enable

Add to ~/.mcpproxy/mcp_config.json (all optional — defaults already spotlight untrusted output):

{
  "output_sanitisation": {
    "spotlight_untrusted": true,
    "response_action": "redact",
    "strip_control_chars": true,
    "strip_classes": ["ansi", "c0c1", "bidi", "zero_width"],
    "max_redactions": 100
  }
}
  • response_action: "spotlight" (non-mutating) | "redact" (mask secrets) | "block" (replace payload on critical detection).
  • Fully opt-in: omit the block (or leave spotlight_untrusted: false) → mcpproxy forwards everything unchanged. Set spotlight_untrusted: true to enable the wrapper.

Verify (the mandatory three)

  1. curl / MCP roundtrip: stand up mcpproxy with a stub untrusted upstream that returns a secret + ANSI; call the tool via /mcp and call_tool_read; assert the response is wrapped («untrusted:…») and, with redact, the secret is [REDACTED:cloud_credentials].
  2. e2e script: ./scripts/test-api-e2e.sh stays green (no regression).
  3. Web UI / chrome ext: open the activity view; confirm a policy_decision row appears for the redact/block call with the reason.

Behaviour summary

Config Untrusted tool text Trusted tool text Non-text blocks
default wrapped in delimiters unchanged unchanged
redact wrapped + secrets masked secrets masked unchanged
block + critical payload → remediation error + audit payload → remediation error + audit n/a
strip_control_chars wrapped + control seqs stripped unchanged unchanged