OnclaveAgent

Onclave Networks, a B2B cybersecurity firm, had the same leak almost every professional company has: visitors arrive on the website at all hours with real questions — what the product does, how it compares, whether it fits their environment — and a static site can only point them at a contact form. The people most ready to engage often show up outside business hours, and a form they don't trust loses them. The brief was to put a knowledgeable, always-on layer on the site that could actually answer from the company's own material, tell a genuine prospect apart from a sales pitch, and capture the ones worth following up — without ever bluffing on something it shouldn't answer.

We built OnclaveAgent, a retrieval-augmented (RAG) chat agent that answers visitor questions strictly from Onclave's real, approved content, with citations back to the source. Rather than a generic chatbot, it runs explicit four-case conversation routing: answer with a citation when it knows, ask one clarifying question when the request is ambiguous, say plainly "that's not in our materials" and take a message for anything it shouldn't quote, and politely redirect off-topic chatter. A core design rule — lead with the limit, never bury it — means the agent surfaces what it cannot do instead of guessing. When a visitor signals real intent, it classifies the conversation (pricing, demo, contact, sales) and pops a lead-capture form, then routes the captured lead to the team.

OnclaveAgent is built in TypeScript on Next.js with the Claude API as the reasoning layer over a retrieval index of the company's approved content. Each answer is grounded in retrieved passages and returns citations, so responses are traceable rather than free-generated. An intent classifier tags every conversation, driving both the lead-capture trigger and a back-office analytics dashboard that shows the top questions asked, the leads captured, and the gaps where visitors wanted something the site doesn't yet explain. Before launch — and as a guard against regressions — the agent runs against a 20-question evaluation suite covering both retrieval accuracy and chat routing, with a measured baseline, so changes are tested against real questions rather than shipped on vibes.

OnclaveAgent runs live and public on Onclave's site today, answering visitor questions from real content, routing by intent, and capturing leads around the clock. Because it's showable, it doubles as the clearest possible sales proof for Kaydenlabs: rather than pitching the concept of an AI agent, prospects can watch a real one answer, refuse appropriately, and capture a lead on screen. It is the same engine that powers the Kaydenlabs AI Front Desk for professional firms.