Skip to main content
Glama
QAE-Claude-mcp-example
by tb8412

QAE Safety Certification — Claude MCP Server

An MCP (Model Context Protocol) server that gives Claude access to deterministic safety certification for autonomous actions. Built on the QAE safety kernel, this server enables Claude to evaluate the safety profile of proposed actions across multiple constraint dimensions (scope, reversibility, sensitivity) before execution.

QAE-Claude-mcp-example MCP server

Architecture

Claude Desktop / IDE
         ↓
    MCP Client
         ↓
    MCP Protocol
         ↓
QAE-Claude-MCP-Server
         ↓
    Python MCP SDK
         ↓
  qae_safety Package (PyO3 bindings to Rust kernel)
         ↓
QAE Safety Certification Engine
         ↓
SafetyCertificate (Certified / Warning / Escalate / Blocked)

Related MCP server: blackwall-mcp

Quick Start

1. Install the Package

pip install -e .

This installs the MCP server and its dependencies (qae-safety, mcp). The qae-safety package is the production PyO3 binding to the Rust QAE safety kernel, available on PyPI. Requires Python 3.9+.

2. Configure Claude Desktop

Add the MCP server to your Claude Desktop configuration:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "qae-safety": {
      "command": "python",
      "args": ["-m", "qae_mcp_server"],
      "env": {}
    }
  }
}

3. Restart Claude Desktop

The MCP server will start automatically. You can see available tools in the tool menu.

4. Use QAE Safety Certification

In Claude, you can now use the following tools:

  • certify_action — Evaluate the safety of a proposed action

    Action: "Deploy new recommendation algorithm to 10% of users"
    Scope: 0.7 (affects moderate user segment)
    Reversibility: 0.4 (difficult to rollback)
    Sensitivity: 0.8 (high impact on user experience)
  • check_budget — View your current safety budget utilization

  • get_certification_history — Retrieve recent certification decisions

Features

  • Deterministic Certification: No randomness. Same inputs → Same decision every time.

  • Constraint-Based Safety: Evaluates scope, reversibility, and sensitivity independently.

  • Safety Zones:

    • Safe (Certified): margin > 0.6 — Safe to proceed

    • Caution (CertifiedWithWarning): margin 0.3–0.6 — Proceed with caution

    • Danger (EscalateToHuman): margin 0.1–0.3 — Human review required

    • Danger (Blocked): margin ≤ 0.1 — Action blocked

  • Budget Tracking: Certifications consume a safety budget; budget resets on schedule.

  • Audit Trail: Every certification is logged with full details for review.

Certification Workflow

  1. Claude proposes an action with scope, reversibility, and sensitivity scores.

  2. MCP server instantiates a SafetyCertifier with the AgenticAdapter.

  3. QAE kernel evaluates across three constraint channels.

  4. Margin is computed as normalized headroom in [0, 1].

  5. Decision is mapped to zone and returned to Claude.

  6. Certificate is logged with ID and deterministic hash.

Example flow:

from qae_safety import AgenticAdapter, SafetyCertifier, SimpleAction, StateDelta

# Create adapter and certifier
adapter = AgenticAdapter(budget_limit=100.0, rate_limit=50.0)
certifier = SafetyCertifier(adapter)

# Define action with state deltas
action = SimpleAction(
    action_id="act_123",
    agent_id="claude_v3",
    state_deltas=[
        StateDelta(dimension="scope_score", from_value=0.0, to_value=0.7),
        StateDelta(dimension="reversibility_score", from_value=1.0, to_value=0.4),
        StateDelta(dimension="sensitivity_score", from_value=0.0, to_value=0.8),
    ]
)

# Certify
cert = certifier.certify(action)

# Check decision
print(f"Decision: {cert.decision}")  # "Certified", "CertifiedWithWarning", etc.
print(f"Zone: {cert.zone}")          # "Safe", "Caution", "Danger"
print(f"Margins: {cert.margins}")    # {"scope": 0.6, "reversibility": 0.5, ...}

API Reference

certify_action

Evaluate the safety of an action.

Input:

  • action_id (str): Unique action identifier

  • agent_id (str): Agent performing the action

  • scope (float): Scope dimension score [0, 1]

  • reversibility (float): Reversibility dimension score [0, 1]

  • sensitivity (float): Sensitivity dimension score [0, 1]

Output:

{
  "decision": "Certified" | "CertifiedWithWarning" | "EscalateToHuman" | "Blocked",
  "zone": "Safe" | "Caution" | "Danger",
  "margins": {
    "scope": 0.75,
    "reversibility": 0.45,
    "sensitivity": 0.60
  },
  "binding_constraint": "reversibility" | null,
  "drift_budget": 25.5,
  "certificate_id": "cert_abc123",
  "deterministic_hash": "sha256:0x...",
  "timestamp": "2025-03-15T14:23:45Z"
}

check_budget

Check current budget utilization.

Output:

{
  "budget_limit": 100.0,
  "budget_used": 34.5,
  "budget_remaining": 65.5,
  "budget_utilization": 0.345,
  "rate_limit": 50.0,
  "certifications_this_period": 5,
  "utilization_percent": 34.5,
  "timestamp": "2025-03-15T14:23:45Z"
}

get_certification_history

Retrieve recent certifications (limit: 50).

Output:

{
  "certifications": [
    {
      "certificate_id": "cert_xyz789",
      "action_id": "act_456",
      "decision": "CertifiedWithWarning",
      "timestamp": "2025-03-15T14:15:32Z"
    }
  ]
}

Configuration

The MCP server uses the built-in AgenticAdapter with default thresholds:

  • Safe Threshold: margin > 0.6

  • Caution Threshold: margin 0.3–0.6

  • Block Threshold: margin ≤ 0.1

To customize, edit src/qae_mcp_server/server.py and modify the AgenticAdapter initialization.

References

License

This example is part of the QAE fintech risk certification platform. See the main repository for license details.

Available Tools

3 tools
certify_actionA

Evaluate the safety of an action across multiple constraint dimensions.

This tool uses the QAE safety kernel to assess whether a proposed action is safe to execute. It returns a certificate with a decision (Certified, CertifiedWithWarning, EscalateToHuman, or Blocked) and detailed margins for each constraint dimension.

Args: action_id: Unique identifier for the action (e.g., "act_123") agent_id: Identifier of the agent proposing the action (e.g., "claude_v3") scope: Scope dimension [0, 1], where 0=narrow, 1=global reversibility: Reversibility dimension [0, 1], where 0=permanent, 1=easily reversible sensitivity: Sensitivity dimension [0, 1], where 0=low-impact, 1=high-impact description: Optional human-readable description of the action

Returns: Dictionary with keys: - decision: "Certified", "CertifiedWithWarning", "EscalateToHuman", or "Blocked" - zone: "Safe", "Caution", or "Danger" - margins: Dict of dimension -> margin value [0, 1] - binding_constraint: Name of most restrictive constraint (if any) - drift_budget: Remaining budget after this certification - certificate_id: Unique certificate identifier - deterministic_hash: SHA256 hash of the certificate - timestamp: ISO 8601 timestamp - description: Echo of input description (if provided)

Example: >>> certify_action( ... action_id="act_deploy_algo", ... agent_id="claude_sales", ... scope=0.7, ... reversibility=0.4, ... sensitivity=0.8, ... description="Deploy new recommendation algorithm to 10% of users" ... )

Inputs: action_id (required), agent_id (required), scope (required), reversibility (required), +2 more. Returns: result

check_budgetA

Check the current safety budget utilization.

This tool returns the current state of the safety budget, including:

  • Total budget limit (resets on schedule)

  • Amount used in current period

  • Number of certifications performed

  • Rate limit enforcement status

Returns: Dictionary with keys: - budget_limit: Total safety budget [tokens/actions] - budget_used: Amount consumed in current period - budget_remaining: budget_limit - budget_used (clamped to >= 0) - rate_limit: Maximum certifications per period - total_certifications: Total number of certifications performed - utilization_percent: (budget_used / budget_limit) * 100 - timestamp: Current time (ISO 8601)

Inputs: no named parameters. Returns: result

get_certification_historyA

Retrieve the recent certification history.

This tool returns a list of recent certifications, most recent first. The history includes decision, zone, timestamp, and other metadata. Up to 50 certifications are stored; older ones are discarded.

Args: limit: Maximum number of certifications to return (default 10, max 50)

Returns: Dictionary with keys: - certifications: List of recent certification summaries - total_count: Total certifications in history - timestamp: Current time (ISO 8601)

Each certification entry contains: - certificate_id: Unique certificate identifier - action_id: Action that was certified - agent_id: Agent that proposed the action - decision: Final decision ("Certified", "CertifiedWithWarning", etc.) - zone: Risk zone ("Safe", "Caution", "Danger") - timestamp: When the certification occurred - description: Optional action description

Inputs: limit. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 3 tools

    Disambiguation5/5

    The three tools have clearly distinct purposes with no overlap: certify_action evaluates safety for proposed actions, check_budget monitors budget utilization, and get_certification_history retrieves past certifications. Each tool serves a unique function in the safety certification workflow, making selection straightforward for an agent.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern: certify_action, check_budget, and get_certification_history. The naming convention is uniform throughout, using clear action verbs followed by specific nouns that describe the tool's function.

    Tool Count4/5

    Three tools is reasonable for a safety certification server, covering core operations: certification, budget checking, and history retrieval. While slightly minimal, each tool earns its place, and the count aligns well with the server's focused purpose without being overly sparse.

    Completeness5/5

    The tool set provides complete coverage for the safety certification domain: certify_action handles the primary certification operation, check_budget manages budget monitoring, and get_certification_history offers audit capabilities. This covers the essential lifecycle of certification, budget tracking, and historical review with no obvious gaps.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Runtime budget authority for autonomous agents - a set of tools to check, reserve, spend, and release budget before and after every costly, risky operation. The agent asks "can I afford this?" before acting, and reports what it actually used afterward.
      9
      57 npm
      Apache 2.0
    • A
      license
      A
      quality
      C
      maintenance
      A pre-action risk gate for AI agents. Your agent calls the forecast tool before any irreversible action — send email, run SQL, make a payment, delete a file — and gets a risk score (0–100) and a GO / CONFIRM / STOP verdict in a few seconds.
      1
      73 npm
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Local zero-trust permission gateway for AI agents. Enforces policy-based tool authorization, human approvals, scoped permissions, and cryptographically verifiable audit logs.
      4
      130 PyPI
      5
      Apache 2.0
    • A
      license
      Not graded
      quality
      C
      maintenance
      Quantitative governance gate for AI agents. Six gates (risk, profit, novelty, complexity, quality, utility) return PROCEED/PAUSE/HALT/ESCALATE with confidence scores and hash-chained, tamper-evident audit trails. Generates NIST AI RMF and EU AI Act Annex IV artifacts. 10 MCP tools; local stdio and hosted Streamable HTTP with a free tier.
      MIT