Skip to content

mcp-server.md

Latest commit

 

History

History
123 lines (85 loc) · 4.2 KB

mcp-server.md

File metadata and controls

123 lines (85 loc) · 4.2 KB
Raw

MCP Server

MultiMind ships an Anthropic Model Context Protocol (MCP) server that exposes the compliance toolkit — PII scanning, redaction, grounding checks, policy gating, and audit logging — as tools any MCP client can call: Claude Desktop, Claude Code, Cursor, or your own agent.

Not to be confused with multimind.mcp, MultiMind's internal Model Composition Protocol for workflow orchestration. The MCP server lives in multimind.mcp_server.

Install and run

pip install 'multimind-sdk[mcp]'
python -m multimind.mcp_server

The server is named multimind-compliance and speaks MCP over stdio. The base multimind.mcp_server package imports without the mcp extra; only running the server requires it.

Client configuration

Claude Desktop (claude_desktop_config.json) or Claude Code (.mcp.json):

{
  "mcpServers": {
    "multimind-compliance": {
      "command": "python",
      "args": ["-m", "multimind.mcp_server"]
    }
  }
}

Claude Code one-liner:

claude mcp add multimind-compliance -- python -m multimind.mcp_server

Environment:

Variable Default Purpose
MULTIMIND_AUDIT_LOG ./multimind_audit.jsonl Destination JSONL file for the audit_log tool

Tools

scan_text(text)

Detect PII in text. Returns finding types and character spans plus per-type counts — never the raw matched values.

{"text": "contact jane.doe@example.com"}
{"findings": [{"type": "email", "start": 8, "end": 28}], "counts": {"email": 1}, "total": 1}

Detected types: email, phone, ssn, credit_card (Luhn-validated), ip_address, iban, passport, dob, api_key (entropy-filtered).

redact_text(text, strategy="mask")

Redact detected PII. Strategies: mask ([EMAIL]), hash ([EMAIL:1f9a3c2b], a stable 8-char tag for correlation without exposure), remove (delete outright).

{"text": "mail jane.doe@example.com now", "strategy": "mask"}
{"redacted": "mail [EMAIL] now", "strategy": "mask", "counts": {"email": 1}, "total": 1}

check_grounding(response, sources)

Score how well each sentence of a response is supported by the given source texts. Deterministic n-gram containment — offline, no LLM call. Catches fabricated sentences; being lexical, it cannot catch paraphrased contradictions.

{"response": "The sky is blue. Elephants can fly.", "sources": ["The sky is blue. Grass is green."]}
{
  "score": 0.5,
  "unsupported_count": 1,
  "summary": "grounding score 0.50 over 2 sentence(s): 1 supported, 0 partial, 1 unsupported",
  "sentences": [
    {"sentence": "The sky is blue.", "verdict": "supported", "score": 1.0, "best_source_snippet": "The sky is blue."},
    {"sentence": "Elephants can fly.", "verdict": "unsupported", "score": 0.0, "best_source_snippet": null}
  ]
}

audit_log(event, metadata?)

Append a caller-declared compliance event to the JSONL audit trail (MULTIMIND_AUDIT_LOG). The event text and any string metadata values are PII-masked before writing, so raw PII never lands in the log; detected types are recorded as counts.

{"event": "user pasted jane.doe@example.com into the chat", "metadata": {"session": "abc123"}}
{"ok": true, "path": "multimind_audit.jsonl"}

Written line:

{"timestamp": "2026-07-06T12:00:00+00:00", "source": "mcp", "event": "user pasted [EMAIL] into the chat", "pii_types": {"email": 1}, "metadata": {"session": "abc123"}}

check_policy(text, block_on)

Pre-send gate: call before pushing content to an external service. Returns whether the text is free of the blocked PII types (case-insensitive type names).

{"text": "ssn is 123-45-6789", "block_on": ["ssn", "credit_card"]}
{"allowed": false, "violations": [{"type": "ssn", "count": 1}]}

Positioning

These are agent-invocable compliance tools: the agent decides when to scan, redact, or gate. For transparent interception of all traffic — every prompt and response, no cooperation required — put the compliance guard in the request path instead with multimind serve (see compliance.md), or wrap a model in-process with multimind.compliance.guard.