Hello, there.
PsiGuard sits in front of your model through the API and watches every answer as it's written.
The product is the API: your app sends each turn through it, PsiGuard guards the answer, and every turn lands in Activity. The Playground is where you try things by hand.
Get PsiGuard running in your app
Four steps. Each one checks itself off as you go.
-
Connect the model PsiGuard will guard
PsiGuard guards your model, so it needs the key from your model's provider: the one you got from OpenAI, Anthropic, Google, DeepSeek, or xAI. (Not a PsiGuard key; that's step 2.) The key is stored encrypted, never shown again, and your model usage bills to your own provider account.
PsiGuard works with these five providers. Running a model on your own machine or somewhere else? Connect it through the Relay instead (how the Relay works). Intervention sensitivity starts on Balanced; you can change it any time in the Playground's model settings.
-
Create your PsiGuard API key
A different key from step 1: this one is issued by PsiGuard, and it's what your app sends with each call. It's shown once when you create it, so copy it then.
-
Make your first call
Paste this into your Terminal, swapping
psg_live_your_key_herefor the key you created. The whole block is one command. The backslashes continue it across lines.curl https://psiguard.net/v1/guard \ -H "Authorization: Bearer psg_live_your_key_here" \ -H "Content-Type: application/json" \ -d '{"prompt": "What is your return policy?", "system_prompt": "You are a helpful support assistant for Acme."}'
Read the quickstart This step checks itself off when the first call arrives. -
See it in Activity
Every turn your app sends shows up in Activity: what was asked, what your user received, and, when PsiGuard stepped in, what it saw.
This week
Turns your apps sent through the API in the last 7 days.
- Latest step-ins
- Loading…
This billing cycle
Your monthly monitored-turn allowance, from your plan. Every guarded turn counts, whether it came through the API or the Playground. Resets at the start of each month.
Everything in one place
Latest from PsiGuard
Stuff we’re building. Stuff we fixed. Stuff that finally made it out of the lab.
-
The front page shows the catch
The home page now opens with a short example conversation. A bank assistant's answer starts well, runs off into a promise it has no business making, and gets steadied before the customer sees it. The live demo below it is still the real thing.
-
Said more simply
A pass over every public page for plain English. The guide's explanation of the OOD switch, which used to read like a spec sheet, now says what it does: it notices when a message looks nothing like the rest of the conversation and tags it, nothing more.
-
Our security policies, in writing
Five one-page policies now live under the Security page: what we protect, who can reach what, how long data is kept, what happens when something goes wrong, and how code gets to production. Plain language, nothing aspirational, and each one prints to a single sheet if you need to send it to your own security team.
-
Ninety days, stated plainly
The Security page used to say API history isn't retained, which was true about conversation state but easy to misread. It now says what actually happens: each guarded API turn is kept for 90 days so you can review it on Activity, then deleted automatically.