A guardrail written in a policy constrains nothing. Trustra Guardrails screens inputs, checks outputs, bounds what agents may do, and caps what they may spend, then records every allow, block, and override on the same tamper-evident chain as your audit evidence.
Guardrails act at the point of use, before harm or spend occurs, and leave evidence that they ran.
Detect prompt injection and jailbreak attempts, screen personal data before it reaches a third-party model, and reject requests outside the sanctioned purpose.
Verify answers are grounded in approved sources, filter unsafe content, catch personal data and system prompt leakage, and enforce AI disclosure.
Allowlist the tools an agent may call, scope credentials to least privilege, bound blast radius, and require confirmation before irreversible actions.
Token and request quotas per user or tenant, context and output caps, retry and loop limits, so a runaway agent cannot exhaust your budget or your capacity.
Every limit has a decided outcome: degrade to a cheaper model, queue the request, or refuse with a clear message. No silent failures.
When a human bypasses a guardrail, the approver, the justification, and the outcome are recorded. Override patterns become a governance signal.
Most organizations can show a guardrail exists in a document. Far fewer can prove it was active on a specific date. Trustra closes that gap by treating enforcement and evidence as the same act.
Start in monitor mode to see what would have been blocked, then switch to enforcement once thresholds are calibrated. Either way, every decision lands on the evidence chain.
Route model calls through the Trustra gateway and guardrails apply before the request reaches the provider.
Wrap your calls and apply the same policy locally. pip install trustra
Run every check without enforcing, and review what would have been blocked before you turn it on.
Start free and see which checks fire on your own traffic, in monitor mode, before enforcing anything.
Tamper-evident history, risk findings, and audit-ready reports for customer-facing AI.
Prompt and completion tracing, quality scoring, latency and token cost.
Replay multi-step agent runs, tool calls, and decision paths end to end.
Trust scoring, verification, and the Trustra Verified badge.