Block + alert
hard stopLive API keys, passwords, payment card numbers — the request never leaves the building. Your agent gets a clean refusal it can handle, and you get the alert: webhook, Slack, email — your pick.
The guardian layer — new
Catch the mistake before it leaves.
Agents are powerful interns with zero instincts. apiegg stands at the door — inspecting every call for secrets, bulk data grabs and runaway egress before they leave the building.
See it work
Four calls walk up.
Four calls walk up.
Watch what happens.
A live simulation — it cycles on its own, or click a scenario yourself.
The guardian gate
interactive demoThe gate inspects every call mid-flight.
The action ladder
Every suspicious call
Every suspicious call
gets a judgement.
Not everything deserves the same answer. The guardian escalates in proportion to the risk.
Ask a human
human-in-the-loopSomething ambiguous? The call pauses mid-flight — approve once, approve always for this project, or deny. Human-in-the-loop at the gateway, where it belongs: not buried in your agent's logs.
Clean & send
tokenizedSensitive values are swapped for [TOKEN_n] before the model ever sees them — the model answers, the data never leaves. Real values are rehydrated in the response, on your side only.
Redirect
re-routedBlock is dumb. Redirect is smart. Sensitive traffic re-routes to your cleared or local models — the work still gets done, just inside your walls. Same endpoint, different door.
The tripwires
What it watches for.
The ways agents actually leak — and what the guardian does about each one.
Secrets & credentials
Keys, tokens, passwords pasted into prompts or tool results
→ block
Bulk PII / regulated data
Customer tables, health records, IDs travelling in bulk
→ redact / redirect
Odd egress
"Why is it sending the WHOLE customer table?!"
→ ask
Injection signatures
Prompt-injection patterns hiding in tool results and web content
→ flag
Custom detectors
Your project codenames and internal terms — learned over time
→ your rule
Shadow mode: watch-only, first
Run it watch-only for a week. See what WOULD have been blocked. Then decide. Zero risk, zero interruptions — just the receipts. When you're satisfied the judgement is good, you flip the dial from "watch" to "guard".
📐 Honest scope
The guardian focuses on secrets, exfiltration and bulk PII — the ways agents actually leak. It's not a content moderator.
False-positive controls are first-class: every detector is tunable, every block reversible, every decision logged and exportable — timestamped records shaped for human review and upcoming AI-oversight rules.
Good questions
Before you ask the door.
The call never leaves. Your agent receives a clean, structured refusal it can handle gracefully — not a crash. You get the alert through your channel of choice (webhook, Slack, email), and the event lands in your audit log. If you'd run shadow mode first, you'd have a report instead.
No. Checks run inline with routing — inside the same milliseconds the router already spends analyzing each request. You get gate-level inspection at router-level speed.
That's the point of a gateway: your agents talk to one endpoint, and the door is the endpoint. As long as apiegg is the door, nothing sensitive walks past the guardian. There's no side entrance to forget to lock.
No. ⚖ The guardian is a safety control and an evidence machine — it gives you controls and records. What those mean for your specific regulations is a question for your counsel, not your router.
Early access
Bad calls stop
Bad calls stop
at the door.
Point your agents at one endpoint that routes the smart calls — and stops the stupid ones. Free during early access.
✨ Free during early access · no credit card
⚖ The guardian is a safety control, not legal or compliance advice. Detector coverage and thresholds are tunable per project — false-positive control is a first-class feature.