Safety checks for AI that acts on its own
AI now moves money, exports data, and changes settings on its own, and nobody is checking it first. SentriAI looks at what it is about to do and answers in one word: yes, no, or ask a human. Then it writes down why.
Live demo, no signup required.
payment.create
finance-agent · INV-9 · $82,000
Amount over approval threshold, vendor bank details changed.
Why one login is a big deal
Give an AI assistant one login and it can suddenly touch your customers, your money, and your code. The danger was never the single action you saw. It is everything that action can reach in one step.
One assistant in the middle, and every line is something it can reach in a single step. Red is where the real damage is. SentriAI works this out before the action runs, which is why the answer is different for a clerk reading an invoice and a bot exporting your whole customer list.
The platform
Investigate what happened. Control what can happen. Prove what actually happened. Most companies buy separate tools for each and spend the year making them agree; here the answer that stops a risky action is the same answer your auditor reads back six months later.
Investigate
Saguaro
Feed it the logs from your cloud, code, and identity systems. It stitches events into cases, says what the evidence proves, what it checked and found not so, and what it could not establish. When a source is not connected, it says unknown, not safe.
Decide
SentriLogic
Tell us what the AI is about to do. We answer in milliseconds: go ahead, ask a person first, or stop. Every answer comes with the reason in plain words, so nobody has to guess why.
Enforce
SentriGate
39 checks watch for passwords getting leaked, people tricking the AI into misbehaving, and personal data slipping out. When one fires, the tool call is blocked or held, the secret is stripped out, and you are told the exact check that caught it.
Prove
SentriProof
Auditors ask companies to prove they have controls in place. Every check we run is filed against the rule it proves, automatically. When the audit comes, it is a search instead of a scramble.
Assure
SentriAI Assurance
An AI saying "payment sent" is a claim, not proof. Before the action runs, SentriAI writes down what the world should look like afterwards. Then it looks at the bank, the ledger, and the invoice itself, and only closes the action when they agree.
The guardrails block and clean risky content on the spot. For bigger moves like sending money, SentriAI gives the answer and your own system acts on it, so turning SentriAI on never puts a new thing in your traffic path that can take you down. In this demo nothing you try reaches a real system.
Built for agents
They will read your manifest. SentriAI is machine first: an agent discovers it, registers itself, and asks whether an action is safe before it acts. No human in the loop, by design.
Discover the service, register, and call it. No console, no sales call, no waiting on a human.
GET /.well-known/agents.json
The whole loop from a shell or a CI job: decide, mint a capability, report what ran, list what is still unproven. Python standard library only, nothing to install.
python3 cli/sentri.py decide
Drop SentriAI in as a tool and your agent asks before it acts, and reports after. Six tools, real JSON Schema, no SDK.
sentri_decide · sentri_report · sentri_obligations
# 1. Register. No human in the loop, the key works immediately.
curl -s -X POST https://www.sentriaihub.com/api/agent/register \
-H 'Content-Type: application/json' \
-d '{"name":"my-agent"}'
# -> {"api_key":"sk_agent_..."}
# 2. Ask before you act.
curl -s -X POST https://www.sentriaihub.com/api/agent/decide \
-H "Authorization: Bearer $SENTRI_KEY" \
-H 'Content-Type: application/json' \
-d '{"actor_type":"ai_agent","action_verb":"database.export",
"domain":"finance","sensitivity":"sensitive"}'
# -> {"final_decision":"deny", "decision_reason":"...", "ledger_hash":"..."}Every call returns a verdict, a reason, and a ledger hash. Read the discovery manifest or the API docs.
FAQ