A lightweight SDK places Guardians next to your agents — your data, policies and model choice stay inside your perimeter. Nothing is sent out to be scored.
One wrapper attaches the full inference-path layer
pip install ccai-guardian # pure-Python wheel, ~4 MB from ccai_guardian import Guardian g = Guardian(ruleset="factset-global-markets") # signed rule-set with provenance agent = g.wrap(agent) # attaches Fabric Guardians G1–G5 + the AGM monitor
Where g.wrap() inserts a Guardian into the agent fabric
User Query
User → Model — Prompt injection, jailbreak
Context Window
Context assembly — Context poisoning, prompt override
Tools & MCP
Model → MCP / tools — Excessive agency, tool misuse
Retrieval
RAG → Model — Embedding manipulation
Output
Model → Output — Sensitive disclosure
Start in observe mode to see how agents reason; move to active control when ready
FX Risk Copilot
AGT-01 · FX Trading · 5/5 points wired
Compliance Sentinel
AGT-02 · Risk & Compliance · 5/5 points wired
FX Algo Trading Agent
AGT-05 · FX Trading · 5/5 points wired
ESG Disclosure Drafter
AGT-06 · Sustainability · 4/5 points wired
Trade Surveillance Watcher
AGT-07 · Risk & Compliance · 5/5 points wired
Liquidity Stress Modeler
AGT-08 · Treasury · 3/5 points wired