A review of one agent or one workflow before it touches real customers, money or data. We check what your AI can actually do against what it should do, run fixed-rule checks, rank every finding by severity, and give you a written verdict: Go, Go with changes, or Not yet. It takes 10 business days and costs $1,200 incl. GST.
Your own AI can write a good plan and a decent checklist. What it cannot do is check itself: it shares your blind spots, it reviews what it is told rather than what is actually running, and it tends to agree with whoever is asking. A review is done by someone who didn’t build it, against the system that is actually running, and we will say "not yet" when that is the honest answer.
No. Every review comes with a plain-English summary for whoever makes the decisions and a technical appendix for whoever builds and maintains the system. Built it yourself or never opened a terminal, either works.
The Scope Check and the Role Card are free. The Agent Readiness Review is $1,200 incl. GST, a set price. Build, Secure and Monitor are quoted from what the review finds, because their size genuinely varies. See Pricing for the full ladder.
No, and nobody honestly can. What we do guarantee is the work, the checks and an honest verdict, including "not yet" when that is the truth. If, within 48 hours of getting access to your project, we find we cannot help, you get back everything you have paid.
Understand is the free way in: the Scope Check and the Role Card, for working out whether handing a job to AI makes sense. Build is making that real: custom work wired into the tools you already use, not a template. Secure is checking that what got built only does what it's supposed to, whether we built it or someone else did. Monitor is a scheduled re-check once it's live, against its Role Card and the review findings. Every engagement starts with the same free Scope Check, regardless of which one fits.
No. Virasai works with startups and small businesses, at any technical level: a free Scope Check and Role Card, a set-price Agent Readiness Review, then Build, Secure, and Monitor. None of it is a subscription product. Sentinel and Magus OpenSecMCP are separate, free, open-source tools we also maintain. Magus Local and Magus Agent API are governance architectures published by VaHive Systems Lab, fully specified but with no public implementation yet, and the pages say so.
Yes, in two ways. The free Role Card walks you through defining one job for AI yourself: what it does, who owns it, what it may touch and what it must never do. And every Readiness Review starts from that definition before checking the real system against it. Neither commits you to buying a build afterwards.
Yes. That is Build: custom harnesses, workflows, MCP and API connections, and integrations, from initial concept through refactoring an existing system. It is quoted from what a Readiness Review finds, the same way Secure and Monitor are.
The same way it works on anything we build ourselves. Secure checks what a system is actually allowed to do against what it's actually doing, using fixed rules rather than trusting anyone's word for it, including ours. Who built the thing being checked does not change the process.
Monitor is a scheduled re-check of an agent that has already had a Readiness Review. On each review date we check it against its Role Card and the review findings: what it can reach, whether approvals are still on, whether anything on its "never" list happened, and what has changed. You get a one-page report and a verdict. We install nothing, run nothing on your behalf, and keep no standing access. It is not live monitoring, and there are no alerts between checks.
Yes, and that includes solo founders, not just teams. The free Scope Check and any paid work that follows are open to startups and small businesses, and none of it requires running anything of ours. Secure covers the Readiness Review and the Governance & Drift Audit, with a Pressure Test as a follow-on; Build and Monitor are open the same way. Assessment limits are stated before an engagement starts.
Questions about VaHive Systems Lab’s papers, the Magus architectures, Sentinel, and the Watch page. None of these are needed to work with us.
No. They are preprints describing proposed frameworks and architectures, published with their own open issues. They should be read as research, not as certification or as evidence that something is deployed.
One governance architecture, specified across two deployment pathways, because where the model runs changes what can be enforced. Magus Local is for a model running on your own hardware; Magus Agent API is for a model accessed as a remote service. Both are fully specified across eight documents each, both are unimplemented, and both pages say so.
A tool that records what a published npm MCP server declares, its artifact digest, file inventory, dependencies, install scripts and entrypoints, and tells you what changed when a new version ships. It reads the exact artifact without running any of it. The same repository includes Watch, a self-hosted monitor that polls the packages you choose and raises a reviewable notice.
One command, no account and no signup: npx magus-mcpsentinel analyze npm <package>@<version> --evidence-dir ./evidence. It prints a schema-valid report and keeps the registry metadata and the exact artifact it read. Nothing from the package under analysis is installed, and nothing in it runs.
Yes, permanently. The analyzer, the CLI, the diff engine and the Watch monitor are all in the public repository under Apache-2.0. There is no withheld tier and nothing held back to sell later. Use is governed by the licence and notices in the repository.
No, and it will not pretend to. It produces evidence about what an artifact declares; you set the policy. Deciding whether a change is acceptable needs to know what the package is for, and that is knowledge you have and a tool does not.
Artifact digest, file inventory, dependencies, install scripts and entrypoints are recorded completely, on every package, every time. If something you approved has moved, you will know. Reading the tool surface out of compiled JavaScript is inference and is published as a lower bound: across a pinned corpus of 50 real published MCP servers, 37 yield a usable tool inventory and only 12 can be resolved completely. Both numbers are stated because only the second permits the conclusion that a tool was removed.
None of the three. Sentinel reads public npm artifacts by name and version, it never accepts credentials, never opens an MCP session, and never inspects live traffic. Watch is single-tenant: you deploy it, you hold the data, and nothing passes through us. The only thing any hosted version would need to know is which public packages you care about.
Our own instance, running against public MCP packages we do not control, published as it stands rather than curated. Every change it has recorded is there with the evidence behind it, including the releases where extraction was incomplete and it says so. It is the same code as the repository, and it is the honest version of a demo.
Anything not answered here: Pricing has the full price ladder, Watch shows the tool working on packages we do not control, and the repository documentation is the operational source of truth for both. For anything else, get in touch.