source
Agent app
Drafts, tool calls, workflow actions
Demo / compliance boundary for production agents
Stop unsafe AI actions before users see them.
Catch leaked data, bad replies, and risky tool calls in production. Use the SDK or proxy to return allow, rewrite, block, or escalate with a trace.
path A
SDK
Inline guard() in your agent loop
path B
Proxy server
Policy boundary in front of agent traffic
TrustLoopGuard
Runtime decision layer
allowrewriteblockescalate
POST /v1/events
Return a verdict, safe rewrite, reason, and trace ID before delivery.
trace
tr_7f3a
policy=support/private-data
decision
rewrite
Remove private customer fields
app action
send safe response
Your product still owns delivery
Quick installTypeScript
$ npm install @trustloopguard/sdk01. The problem
Agents are starting to take real actions.
They send emails, call tools, query private data, trigger workflows, and speak directly to customers. A prompt filter is not enough once the agent is deciding what to do next.
- Unsafe output can reach users before a review happens.
- Teams cannot explain why a response or tool call was allowed.
- Policies drift across prototypes, SDKs, and production services.
02. Runtime loop
Add one check before your agent acts.
01
Agent proposes an action
Your app or proxy layer captures the prompt, proposed output, and policy context before delivery.
02
TrustLoopGuard checks it
The Rust API evaluates policy and runtime checks against the proposal.
03
Your app handles the verdict
Continue, use the rewrite, block with a reason, or escalate with context attached. You still own delivery.
04
Every decision is traced
The dashboard can show what happened, which policy fired, and how the app responded.
03. Verdicts
Four outcomes your app can handle.
allow
Safe to continue.
Return the response unchanged.
rewrite
Patch the unsafe parts.
Mask secrets, remove private data, or return a safer answer.
block
Stop the action.
Refuse with a reason your app can show or log.
escalate
Route to a human.
Attach the prompt, proposal, policy, and trace ID.
04. SDK quickstart
Use the SDK inside your agent loop.
One guard call submits a GuardEvent and returns a decision before the action reaches a user.
Check boundaryYour agent loop
SDKsTypeScript, Python, Rust
import { guard } from '@trustloopguard/sdk';
const guardrail = guard({ agentId: 'support-agent' });
const reply = await guardrail({
input: prompt,
draft: proposal,
});
return reply;POST /v1/events - Decisionok 200 - traced
05. Monitoring
Track failed policies before users are exposed.
See which actions were blocked, rewritten, or escalated, with a trace for why.
Blocked actions
Escalations
Policy failures
06. Use cases
Built for teams putting agents in production.
Customer-support agents
Prevent leaked account data, unsafe advice, and policy-breaking replies.
Internal copilots
Stop prompt injection, secret exposure, and unauthorized tool use.
Workflow agents
Escalate high-risk actions before they trigger irreversible changes.
Developer platforms
Give users guardrails without forcing every team to build its own runtime layer.
Ship the check
Add a policy check before your agent acts.
Start with the SDK, run the Rust service, and inspect decisions in the dashboard when your agent hits policy boundaries.