Your AI agent can move money, send email, run code — on its own. VENZX pauses high-risk actions for a human, keeps a tamper-evident record of every one, and blocks what breaks your rules.
500 free checks·No card required·No subscription, no lock-in
Three outcomes
No model in the path. Every decision is a plain, deterministic rule — returned in single-digit milliseconds, with a reason you can audit.
Nothing violates your policy. The action runs.
It breaks a rule — off the allowlist, an internal address, a leaked key. Stopped before it runs.
Permitted but high-risk. Paused until a human approves.
A real verdict, in ~8ms
“send_email” needs human approval before it runs. The agent is paused and you’ve been pinged — it won’t send until someone says yes.
Sits in front of the tools you already build with
Each entry is sealed to the one before it. Edit or delete anything after the fact and the chain breaks — visibly, permanently.
The more an agent can do on its own, the more there is to go wrong. None of these are hypotheticals — they're happening in production right now.
A stuck agent loop can run up real money fast — one developer's hit $26 in a single turn. With no ceiling, nothing stops it from doing it again.
A hidden instruction buried in a document or email is all it takes for an agent to fire off an email, hit an API, or delete something — no one asked, no one approved.
When a customer asks “how do I know your AI didn't do something it shouldn't?”, a model gives you nothing to show. That silence can cost you the deal.
Say which tools run freely, which need a human first, and a hard spend cap. VENZX decides before every action — and you get a signed receipt of each one.
$ pip install venzxfrom venzx import Venzx, Policy vx = Venzx() guard = vx.guard_for(policy=Policy( tools_allowlist=["search"], # run freely tools_require_approval=["send_email"], # pause for a human max_tool_calls_per_run=20, # circuit breaker )) guard.tool_call("send_email", {...}) # -> ApprovalRequired: paused
Your agent doesn't talk to the outside world directly. It goes through VENZX first.
It tries to send an email, call an API, or use a tool.
It checks the action against your rules, your spend cap, and your approval list.
Safe actions pass. Dangerous ones are stopped. Risky ones pause for a human to approve.
Every decision is recorded, and you're pinged the moment it blocks.
High-risk actions — send, pay, delete — wait for a person to approve. The approver gets a one-click email; the agent resumes only when they say yes, and auto-rejects if no one does.
You say which tools run freely. VENZX allows the safe ones and blocks the dangerous ones — before they happen.
A hard ceiling on calls and cost per run, so a runaway loop can't quietly burn through your budget.
No subscription. Sign up free, get credits to try it, then top up when you need more. Stop one bad action and it's paid for itself.
Credits are non-refundable except as required by law — see our Terms.
Each check runs in a few milliseconds and happens in line with the action your agent was already taking, so in practice you won't notice it. Measure it on your own workload before you depend on a specific number.
VENZX runs as a hosted API. You send the action or text to be checked, it returns a verdict, and a record of the decision is stored so you can search and export it. Sensitive values are masked in what's stored — the log keeps proof of what happened, not the raw content itself.
VENZX sits at the point where your agent calls a tool or model, so it isn't tied to one framework. It's designed to slot in front of common stacks like LangChain, LlamaIndex, and CrewAI, as well as direct API calls.
The action is held and the agent pauses. An approver gets an email with one-click Approve / Reject buttons — no login needed — and the agent resumes only if they approve. If no one acts, it auto-rejects after a timeout you set (default 2 hours). Every request, decision, and who made it is written to the tamper-evident log, which you can export as CSV or a printable PDF, or watch on a read-only status page.
Signing up is free and includes 500 free checks, no card required — enough to properly try it on a real agent. After that, you load credits whenever you want — for example $10 for 1,000 checks. No subscription, and credits never expire.
Every entry carries a fingerprint of the entry before it, forming a chain. Changing or removing any past entry breaks the chain at that point, which is detectable. It doesn't prevent tampering — it makes tampering impossible to hide.
One line of code. 500 checks free, no card. It fails open by default, so an outage can never take your app down — and you can pull it out just as fast if it's not for you. The only risk is shipping without it.
Free to start·No card required·Set up in one line