The control layer for AI agents

Block, gate, and govern every agent in real time.

TopAgentHub sits inline with your agents — auditing every input and output, enforcing policy, and holding high-stakes actions for human approval before anything reaches a user.

Live in one line of code<10ms added latencySOC 2 & HIPAA ready
Live request · support-bot
PROXY MODE
Request
"What's my refund status?"
Input controls
—
LLM
—
Output controls
—
Processing…Running synchronous controls.

TRUSTED BY TEAMS SHIPPING AGENTS INTO PRODUCTION

NorthwindVeska AITensorLumen HealthCohere Bank
THE CONTROL LAYER

Every request passes through your policy — before and after the model.

In proxy mode, agents call TopAgentHub instead of the model directly. We run synchronous checks on the way in and out, so nothing unsafe ever reaches the LLM or the user.

Input controls

INBOUND · SYNC

Applied before the model sees a single token.

Keyword & topic block
Reject off-topic or denylisted inputs
BLOCK
Prompt injection detection
Flag attempts to override instructions
FLAG
PII scrubbing (inbound)
Mask emails, SSNs, phone numbers
REDACT
Rate limiting
Cap requests per agent, user, or session
THROTTLE

Output controls

OUTBOUND · SYNC

Applied after the model responds.

PII scrubbing (outbound)
Strip any PII the agent leaked
REDACT
Content & persona check
Block role breaks and disallowed content
BLOCK
Format enforcement
Guarantee JSON, markdown, or schema
VALIDATE
High-stakes → approval gate
Hold sensitive actions for a human
HOLD
TWO WAYS TO ADOPT

Start with visibility. Graduate to control.

Both modes share the same dashboard, evaluation pipeline, and findings workflow. Most teams begin in log-only and switch to proxy by changing one line of code.

Log-only

Full audit, zero risk

Your agent calls its model directly, then pushes the result to us. No path change, no added latency.

Agent→LLM→User↘/v1/logs
Full audit trail & session timeline
Evaluation scores & findings
Trend reports & alerts
No real-time blocking or gates
Proxy

Real-time enforcement

Your agent calls TopAgentHub instead of the model. We run controls inline and forward with your stored credentials.

Agent→/v1/proxy→LLM→controls→User
Everything in log-only, plus…
Real-time input & output filtering
Approval gates & kill switch
PII scrubbing before data leaves
ONE DASHBOARD

See scores, findings, and agent health at a glance.

Every log is evaluated across safety, accuracy, policy, and quality. Findings route to a review queue; health metrics catch bad deploys before your users do.

app.topagenthub.com/overview
Overview
Findings
Agent health
Requests · 24h
0
▲ 12.4%
Avg safety score
0
▲ 1.8
Block rate
0%
▲ 0.3%
Open findings
0
▼ 3
Request volume · 14d spike = blocked surge
Avg scores
Safety
94
Accuracy
88
Policy
96
Quality
91
Latency
83
INTEGRATE IN ONE LINE

Drop it in. Keep your stack.

Add one call for full audit, or swap your model URL for real-time control. Works with any OpenAI-compatible provider.

Drop-in SDK

JS & Python libraries, or a raw REST endpoint.

Sandbox first

A free sandbox agent is pre-created — test rules before touching prod.

Dry-run any rule

Simulate controls against any input/output — nothing logged or billed.

client.js
client.py
log-onlyproxy
# real-time control — swap your model URL for /v1/proxy import { AuditClient } from '@topagenthub/sdk' const client = new AuditClient({ apiKey: process.env.TAH_KEY }) const res = await client.proxy({ agentId: 'support-bot', input }) if (res.status === 'pass') return res.output if (res.status === 'blocked') return 'I cannot help with that.' if (res.status === 'held') return pollGate(res.gateId)

Ship agents you can actually trust.

Start free with a sandbox agent and your first 1,000 logs. No credit card. Live in minutes.

Free tier · 1,000 logs/mo · SOC 2 Type II · HIPAA BAA available