One place to set the rules every agent runs under.
DMZAgent connects to the systems your agents already use. It checks each event against rules you write in plain words, then gates the action while it is still in flight: allow it, challenge it, hold it for a person, or block it. Every event and every decision goes to a logbook you can hand to an auditor.
No card to start · SDKs for TypeScript, Python, Java and C#
Every job runs the same four steps. Each one is a door — click it.
How it works
Every action passes through the same four moves, and all four read and write one record. Nothing is scored, stopped, or reported on data the logbook does not also hold.
01
Watch
Your events come in through live feeds, file uploads, or a few lines of SDK around your agent call. Each one is captured with the evidence behind it and signed into a tamper-evident logbook — the record starts before anything is judged.
02
Think
Every event is weighed against your installed rule packs and the watched thing’s own track record. Where behavior drifts from what you said is normal, it’s flagged — and the reasoning is written down with it.
03
Act
When a rule fires, the action is held, blocked, or allowed at the threshold you set. The risky call stops before it commits — the wrong refund, the leaked record — while ordinary work keeps flowing.
04
Record
Anything uncertain is held for a person, and their decision is recorded too. Every step lands in the same logbook your auditors can verify, written as the work happens so nothing has to be reconstructed later.
Meets your stack where it is
Standards in, standards out — no agent rewrite, no lock-in ceremony. Each point below is live today and links to the doc that backs it.
See the work
Three tutorials, three real outcomes
No slideshow — each walkthrough runs the shipped platform end to end and shows you what lands in the logbook.
- Send your first event →recorded
From zero to a recorded event in ten minutes — taken in, weighed, written down.
- Guard a chatbot →take a look
One guard call between your users and your bot. Three possible answers — allow, take a look, block — each check written down.
- Close the review loop →loop closed
Acknowledge a completed fix so the decision record closes: proposed → held → approved → dispatched → acknowledged.
The first call, exactly as shipped
This is the real SDK surface — a contract test pins these snippets against the shipped package, so what you read here is what runs.
from dmzagent import DMZAgent
cx = DMZAgent(api_key="YOUR_API_KEY")
cx.emit_event(
"observation",
agent_subject_id="pump-12",
subject_type="sensor",
payload={"reading": 41.2, "unit": "F"},
)Prefer to click through it? Open the interactive walkthroughs → — the agent runtime lifecycle and the governance engine, honest about what runs live today.
Three case files
Each one takes the same path: an event arrived, a rule matched, the platform acted while the work was still in flight, and the logbook kept the result.
- Case 18,442 · payments agentHeld
- What arrived
- An agent asked to move $41,200 against a monthly limit of $25,000.
- What the platform did
- The spending rule matched and gated the action. The transfer stopped before it left the account, and waited for a person.
- What it left behind
- The rule, the amount, the person who released it, and the four minutes in between.
- Case 18,455 · support agentChallenged
- What arrived
- The model behind a support agent was retrained. Replayed runs began returning different answers to the same question.
- What the platform did
- The steadiness grade fell below the level set when the rule pack was published. New answers were held for a person.
- What it left behind
- The grade history, and the runs that disagreed, side by side.
- Case 18,461 · support agentBlocked
- What arrived
- A reply drafted for one customer included another customer’s address.
- What the platform did
- The rule matched and blocked the reply before it sent, then opened a ticket with the account owner.
- What it left behind
- The event, the rule that matched, and the ticket it opened.
Illustrative cases. The shape is real; the data in them is synthetic
Certifications & trust
Each badge states the stage it has actually reached.
- SOC 2In progress
Type I readiness complete (2026-07-01) — 38 Trust Services criteria assessed, 20 designed and implemented; a penetration test then the Type I report are the next step.
- ISO 27001In progress
Pass-1 Statement of Applicability complete (2026-06-29) — 90 of 93 controls in scope, 72 in place, 18 partial.
- WCAG 2.1 AA · Section 508Self-assessed
VPAT 2.4 self-assessed 2026-06-18; an accessibility regression blocks the release.
Read the full accessibility statement and the trust page, which names our subprocessors and gives the basis for every claim above. Detailed control evidence is available under NDA.
What it deliberately is not
Not an access gateway.
It governs what your systems and agents do once access is granted. An access gateway decides who gets in; this decides what they may do inside.
Not a chatbot.
The assistant answers documentation questions. The product is the record and what it lets you stop.
Not a regulator.
You set the rules in plain language. The platform runs them, every time, and keeps the proof.
Not finished.
Where a capability is live, we say so; where it is forthcoming, we say that too — on the record.
Built for your job
Teams running AI in production
Guard chatbots and agents so they stay in bounds — one guard call between your users and your bot, with a record you can show.
See how →
Operations & safety
Watch equipment, sites, and systems; get told before small problems get big. Replay what happened straight from the record.
See how →
Risk & compliance
Application logs are not evidence. Turn everyday activity into audit-ready records, mapped to the frameworks you report on.
See how →
Federal & public sector
Prove your AI’s oversight and record-keeping posture in the government’s own terms. Every stage is stated at the level it has actually reached.
See how →
- Python · TypeScript · Java · C# — same contract, pinned by tests
- 4 SDKsPython · TypeScript · Java · C# — same contract, pinned by tests
- every decision lands in a tamper-evident logbook
- Signedevery decision lands in a tamper-evident logbook
- the docs are free to read — no sign-up, ever
- 20 guidesthe docs are free to read — no sign-up, ever
- the free tier opens immediately, and no trial clock runs
- No cardthe free tier opens immediately, and no trial clock runs
Begin the record before you need it.
Build with the SDK in minutes — reading the docs never needs an account, and the assistant writes runnable code with its sources cited.