The value is the AI extension rather than the checklist. ISO 27001 was written for information systems, and the threats that matter for a model — adversarial evasion, poisoned training data, weights walking out the door — have to be mapped onto Annex A by someone. This does that mapping and produces the artifacts an audit asks for, chiefly the Statement of Applicability and a phased remediation roadmap. It scores what you tell it you have implemented, so it is a second reading of your control list, not evidence.
An MCP server for ISO/IEC 27001:2022 work on AI systems. Annex A's 93 controls span four themes — Organizational (37), People (8), Physical (14) and Technological (34) — and the server scores an ISMS against all of them, then extends into ground the standard does not cover directly: adversarial attacks, model theft, training-data breach and supply-chain compromise as threat scenarios, plus a crosswalk into ISO 42001.
- Per-theme compliance status across all 93 Annex A controls, with gaps identified, critical controls flagged and a certification-readiness verdict — `audit_isms`
- An ISO 27005 risk assessment over 10 AI-relevant threat categories, scored as likelihood against impact, with a treatment plan naming specific Annex A controls — `risk_assessment`
- A gap analysis against a full, core or ai-focused target, returning a prioritised roadmap in three phases: critical, standard and remaining — `gap_analysis`
- A Statement of Applicability per clause 6.1.3(d), documenting every control as implemented, excluded with a justification, or not yet addressed — `generate_soa`
- Incident classification against controls A.5.24–A.5.28: severity, priority, notification requirements and response steps, including AI categories such as data poisoning and prompt injection — `incident_classification`
- A mapping from the ISMS controls you already hold to ISO 42001 AI-governance requirements, showing where AI-specific controls are still needed — `crosswalk_to_ai`
No account and no key. The registry manifest publishes it as the PyPI package `iso-27001-ai-mcp` 1.0.14 over stdio with a Python runtime; the client entries in the README launch it through npx instead. Everything the tools score comes from what you pass in — organization context, scope description, the controls you list as implemented — so the output is a structured reading of your own control list, not an inspection of your systems.
One command — pip install iso_27001_ai_mcp