In private preview
A widget you drop into your site or app. It answers visitors from a knowledge base built out of your pages, documents and source code — citing its sources, in the visitor's language — and hands the rest to your team as tickets.
It either knows, or it escalates. Never invents.
Every answer names the sources it was built from. No knowledge, no answer — never a guess dressed up as one.
What it cannot answer becomes a ticket for your team, with the whole conversation attached.
Hold every answer as a draft for a human to approve, edit or reject — or only the low-confidence ones.
Point Minervia at your site and documents — or run the CLI to scan source code and databases too.
The knowledge base is yours to read and edit in the dashboard. What the assistant knows is never a mystery.
One script line adds the chat to your site — or embed it as a component inside your existing app.
Unanswerable questions become tickets your team works through in the backoffice.
The same knowledge, run two ways — switch on the sales job, the support job, or both. Each is its own agent on your site.
| Sales — the site agent | Support — the support agent | |
|---|---|---|
| The visitor | Anonymous. No account, no sign-up, nothing stored in the browser. | Remembered. One deletable id, written only at first chat use — with a "forget me" control. |
| The conversation | Instant answers to product and site questions, sources cited. | Becomes a numbered ticket with states — open, pending, resolved — the visitor can revisit. |
| Your team | Not needed. | Takes over inside the same ticket; optional supervision holds answers for approval first. |
The ingestion reads it all and rewrites it into one clean, deduplicated knowledge base you can inspect and edit.
Docs go stale; code never lies about current behaviour. The CLI scans your source on your own machine and turns what it finds into support knowledge — so the answer to a ticket can come from the one place that is always right.
One small program you run on your own machine. It builds your agent's knowledge from the inside — and nothing leaves without your approval.
You say where your information lives — source code, folders, databases, your site — and it reads them and writes the articles from what it finds. Every AI call is written to an audit you can read.
minervia source add folder ./src minervia scan You read every article and approve or exclude each one. Only approved articles ever leave your machine; excluded ones stay behind, every time.
minervia articles Sign in once — the browser opens and you approve, no URLs to type. Then one command publishes the approved bundle to your agent: the dashboard shows what arrived, and a re-run only uploads what changed.
minervia link minervia deploy Install it as a service and the CLI becomes your Connector: your agent asks it for live answers — a price, stock, an order's state — over an outbound-only link. No open ports; your credentials never leave your network.
minervia service install One binary, no installer — download and run. The Windows build is signed by Karditor; macOS and Linux builds are unsigned for now.
The pieces you switch on as you need them — piece by piece, or all at once with Private.
The sales-side chat for every visitor — anonymous, instant, sources cited under each answer.
Tickets, states and human delegation — ready to drop into any existing app. Optionally it captures your app's logs and attaches them to the ticket, so the report arrives with the evidence inside.
Scans source code and databases on your machine and drafts knowledge from what it finds — you approve every entry before it leaves.
The CLI installed as a service inside your network. Outbound-only, a human-approved allowlist of read-only tools, and a local ledger of every execution.
Point the assistant at an endpoint of your API — URL, auth and a description it reads. Timeouts, call limits and a domain allowlist bound every call.
Browse every session and ticket, take over conversations, and approve, edit or reject held answers — one queue, sortable by confidence.
Resolved tickets become references: on a new case, your team sees how the last similar one was answered — and what worked.
When conversations expose a gap — a repeated ask, a missing feature — the agent files an improvement request linking the sessions behind it. You accept, reject or merge.
The site agent writes nothing to the visitor's browser. The support agent writes one opaque random id, only at first chat use — never on page load, never across sites.
A control in the widget clears the id and deletes the linked history server-side. Retention is yours to configure.
Run the language model on your own machine through the Connector. Visitor questions and answers never leave your infrastructure.
We run the model on our provider keys. Usage comes off prepaid credits at what the provider charged that day plus the fee — the rates are published in your dashboard — and when the credits run out the agent says so and degrades to tickets.
Plug in your own provider account. You pay your provider for the tokens and there is no token line here at all — the same door, the same ceiling, the same usage record.
Your machine, through the Connector, dialing out. No provider bill and no token line — and no visitor data leaving your network.
The subscription is what you pay us. The AI is passed through at what the provider charged, plus the intermediation fee that covers moving the money and nothing else.
An assistant on a site, with its knowledge base, widget and ticket flow. 25,000 tokens a minute. Plus at $45 and Pro at $100 buy four and twelve times the throughput.
The CLI installed as a service in your network, standing to answer the live queries you approved.
A person who reviews sessions, approves drafts, edits knowledge. Your first one is included.
Opening credit. Enough for an agent and a connector with room for the AI, or two agents with more.
Priced by the month, charged by the day. You buy credits, and a switch in your dashboard turns an agent or a connector on. From then on it costs one day at a time — $20 a month is 67 cents a day — and turning it off stops it that same day. No period, no renewal, no invoice, nothing to cancel. Need an agent on Mondays? Pay four days.
AI is metered by the token, at the provider's own price for that day plus the fee — the per-model rates are in your dashboard and re-read every day. Bring your own key and there is no token line at all. There is no monthly quota: asking more costs us nothing more, so what a plan sells is speed — above its ceiling the agent waits a second rather than refusing. You do set a spend limit per visitor, so no stranger can empty your balance. No setup fees, no seat minimums, no annual lock-in.