AIR Blackbox MCP Server is the MCP server component in the AIR Blackbox ecosystem, used to perform EU AI Act compliance scanning on Python AI agents and to expose auditing, replay, scanning, and compliance tools to any MCP-compatible agent or IDE.
Purpose
This project is aimed at developers using MCP-compatible clients such as Claude Desktop, Claude Code, and Cursor, helping them perform EU AI Act compliance checks, prompt injection detection, risk classification, remediation generation, GDPR compliance checks, bias analysis, and compliance reporting on AI agent code. The frameworks it supports include LangChain, CrewAI, AutoGen, OpenAI, Haystack, LlamaIndex, Semantic Kernel, Google ADK, Claude Agent SDK, and general RAG pipelines.
Main Capabilities
The tools are divided into two tiers:
- Basic version (10 tools, no SDK features): can run independently using only the built-in lightweight scanner, covering scanning of 6 EU AI Act articles and detecting prompt injection (15 patterns).
- Full version (14 tools, including GDPR, bias, validation, and history): requires installing the air-blackbox SDK (>=1.13,<2) to unlock advanced compliance features.
The 4 additional tools unlocked by the full SDK:
- GDPR scan (scan_gdpr): detects personal data processing without consent, data retention and erasure policies, cross-border transfer safeguards, and data processing agreements.
- Bias analysis (scan_bias): detects disparate impact risk, protected attribute handling, training data bias metrics, and fairness metric awareness.
- Action validation (validate_action): pre-execution approval gate (Article 14), ConsentGate policy enforcement, risk-based action filtering, and audit trail generation.
- Compliance history (compliance_history): tracks historical scan results, analyzes compliance trends, exports audit trails, and monitors improvement over time.
The product positioning emphasizes not only scanning but also remediation: discovering issues in 6 EU AI Act articles and generating usable code fixes; combining regex patterns, AI-driven model analysis, and prompt injection detection; validating agent actions before execution through pre-approval gates and risk classification (Article 14); providing GDPR checks, bias analysis, full reports, and historical trend monitoring through the SDK.
Usage
After installation, configure according to the client:
- Claude Desktop: edit ~/Library/Application Support/Claude/claude_desktop_config.json, restart Claude Desktop, and the 14 tools will appear automatically.
- Claude Code / Cursor: add the configuration in the project's .cursor/mcp.json, or in Claude Code's .claude/mcp.json.
After configuration is complete, you can call it directly in natural language, for example: "Scan this code for EU AI Act compliance" "Add a trust layer to this LangChain agent" "Check this text for prompt injection" "What does Article 12 require?" "Generate a compliance report for ~/myproject" "Classify the risk level of send_email" "Scan this code for GDPR issues" (requires the full SDK) "Check for bias in this AI model code" (requires the full SDK) "Can my agent call this shell function?" (requires the full SDK) "Show me my compliance trends" (requires the full SDK)
Optional: if you need AI analysis beyond regex patterns, you can use Ollama for in-depth analysis. Installing [full] unlocks Tier 5 SDK tools; the basic installation works independently.
Regarding MCP SDK compatibility, this package supports two generations of MCP SDK (mcp>=1.0, with no upper limit). mcp 2.0 removes mcp.server.fastmcp and replaces FastMCP with MCPServer; this serviceThe device detects the installed generation and binds to the correct class, so it runs on both 1.x and 2.x. Version 0.2.3 once declared mcp>=1.0.0 with no pinned upper bound, causing newly installed servers to crash on import after mcp 2.0 was released; 0.2.4 fixed this by restricting mcp<2, and 0.3.0 removed the upper bound entirely.
Result Provenance
Every machine-readable result carries a provenance block: engine (builtin-rules or air-blackbox-sdk, the actual runner), ruleset_version (the content hash of the active ruleset, not a manually maintained string), sdk_version (the SDK version that produced the result, null for built-in results). Two reports with the same engine + ruleset_version are generated by byte-identical rules and can be compared directly; different values mean the rules have changed. Error messages also carry provenance.
Applicable Scenarios
Suitable for development and compliance scenarios that require EU AI Act compliance scanning, prompt injection detection, risk classification, remediation generation, GDPR checks, and bias analysis for Python AI agents, and that wish to connect to MCP clients such as Claude Desktop, Claude Code, and Cursor.
Limitations
The basic version does not have GDPR, bias, validation, and history features; these require installing the full air-blackbox SDK. Other specific information such as pricing, charging model, and platform support scope is not yet provided in the materials.
