A from-scratch classroom, like The Kronborg History Comes Alive Project, is real custom development: individually designed agents, a shared emotion and memory engine, live API integrations, and often a full instructor syllabus. This estimates that cost using ETL's actual pricing structure. Every number below is an assumption you can adjust, especially the design hours; treat this as a starting point for a conversation, not a locked price.
Tell it in plain words, not a spec. It'll suggest agent count, tier, and which bespoke systems fit, then you can adjust anything below.
Agents, classroom tier, and the systems this build actually needs.
Default assumes real per-agent QA: personality, voice, portrait/blink, and tool-use checks, rounded up on purpose so a build doesn't get underquoted.
Rate for the specialized design and testing work, not generalist dev or UX rates.
Hosting is metered by real use, every message is a live model call, every bio playback is a live voice call, so it scales with how many students actually show up, not a flat shelf price.
Departments fund builds like this with AI-literacy and education-technology grants more often than you'd think. Before you decide it's out of reach, see what's actually being awarded right now. ETL keeps cost down by retaining the rights to what it builds. Ask about up to 50% off if your project doesn't need exclusivity, some are a great fit for reuse, some aren't, so ask and we'll tell you honestly which yours is.
Why hosting is metered, not flat. Every chat exchange is a live model call and every bio playback is a live voice call; cost tracks real engagement, not time on a shelf. A flat rate either overcharges a small section or quietly loses money on a large one.
Why bespoke work is priced separately from a standard build. A standard agent build (spec, portrait, voice, blink, tool wiring, deployment) is one thing. A shared emotion engine, cross-session memory, multi-agent orchestration, a real instructor syllabus, and custom voice design are systems and specialized work, not agents, and they take real design and testing time beyond any single agent.