Live trust stack
The model is the last thing we trust.
Send a request through Pramagent and watch each guard decide, redact, block, escalate, or allow. Deterministic layers run before the provider is ever called; nothing reaches it until the request passes, and every run is sealed to a verifiable hash chain.
Financial HITL wedge
LangGraph adapter
AutoGen hook
CrewAI guard
OpenAI / Claude / Gemini / Ollama
1Choose a financial safety request
2Or write your own
3Active guards
PII scrubbing
Injection guard
Safety rules
Output judge
Human approval (HITL)
Optional live model key
Leave blank for the deterministic zero-config demo. NVIDIA NIM keys use the selected model; OpenAI keys use gpt-4o-mini; Gemini keys use gemini-2.5-flash. Used only for this request. The demo never persists the key.
4Product signal
Opt in to anonymous demo usage signal
When enabled, Pramagent records only hashed visitor ID, provider kind, verdict, HITL state, and trace-control indicators. No prompts, outputs, API keys, IPs, or plaintext emails.
The trust stack
readySafe output
prev
this
pip install pramagent
github.com/sriram7737/pramagent
Implementation status
The model is not the authority. Deterministic code outside it is.