Summary
The safety framework, in the assistant's reach.
Instead of a document someone reads once, the criteria and their implementation patterns are queryable while the code is being written, including which requirement a pattern answers and what points back at it.
What it is
A Python server that serves the SaferAgenticAI framework to coding assistants: the canonical criteria plus the implementation patterns layer. The package bundles criteria-v1.json, the pattern YAMLs, 4 exemplars and operational_heuristics.yaml, so a wheel install works with no repository checkout.
What you get
- list_suites and list_requirements walk the framework; get_requirement returns one in full
- search_patterns finds the implementation patterns that apply to what you are building
- get_cross_references and get_reverse_references show how a requirement connects, in both directions
- resolve_id turns an identifier into the thing it names
- An editable install picks up pattern YAML edits live, so a team can extend the corpus locally
- Versions can be pinned for an audit trail — the earliest wheels predate the corpus extension and carry fewer patterns with no heuristics, so pin a build from after that landed
Requirements
Python. Install with uvx, pipx, uv tool, pip, or a manual venv from a checkout. No account or key — everything it serves is bundled in the package.
Setup effort
One command — uvx --from git+https://github.com/NellInc/saferagenticai-mcp saferagenticai-mcp