Conversation Simulator

Local-First AI Practice

Pilots get simulators. Surgeons get practice labs. For the conversations that decide careers and friendships, most people get one take — live, unrehearsed, with everything on the line. Conversation Simulator ends that. It’s not a chatbot, not a game, not a course: it’s a practice environment where you run the conversation before it matters, as many times as it takes.

This portfolio page runs the way the product does. Scenario: a skeptic interrogates the studio.

The Portfolio ReviewSkeptical Reader · Turn 1
SKEPTIC · NPC
“Every AI app says it cares about privacy. Why should I believe yours?”
OUTRIGHT MENTAL · YOU
“You shouldn't believe us. There is no server. The language model, speech recognition, voices, transcripts, and scores all run and stay on your machine — and one command proves it.”
Scenario event: receipts_requested triggered
SKEPTIC · NPC
“Fine — show me.”
credibility82
composure71
pressure34

The receipt the skeptic asked for is real, and it ships in the product:

npx convsim offline-smoke-test packs/official/job-interview-basic

One command runs a scripted conversation end-to-end and fails loudly if any subsystem so much as attempts an outbound connection. Trust that can’t be checked is just branding. Check.

What’s under the table

The whole stack is local by architecture, not by policy: llama.cpp runs the character, whisper.cpp hears you, local voices answer, and every transcript and score lives in a SQLite database on your disk. No account. No cloud. No telemetry. No subscription. The conversations you most need to rehearse — the raise, the diagnosis, the apology — are exactly the ones you’d never type into someone else’s server. So there is no server.

Characters aren’t puppets on rails. They push back, warm up, lose patience — live meters show the dynamic shifting as you speak, and scenario events fire when you cross a threshold. When it ends, the debrief tells you what you said clearly, where you hedged, and which moment turned the conversation. That feedback loop is the product.

Written, not engineered

A scenario is a folder of plain YAML — a situation, a character, a scoring rubric, safety rules. No code, no build step. A teacher can script oral exams; a union rep can drill negotiations; a friend can build the exact conversation you’re dreading. Four openly licensed official packs ship with every copy, and the built-in Creator Workbench validates new ones against a published quality bar.

Two rules hold in every scenario, from every author, and no configuration can weaken them: content that sexualizes minors ends the session, and self-harm crisis language ends the session and surfaces real resources. Community packs can make safety stricter. Never looser.

For teams & institutions

Because a scenario is just YAML its author owns, it is also something an institution can commission. A nursing program, a bar association, a social-work faculty, a union local — anyone who trains people through hard conversations can have the exact pack their people need, authored and safety-reviewed by us, then handed over to run on hardware they already control. No seats to rent. No transcripts to leak. The content is a deliverable you keep, not a login you pay for.

SCENARIO PACK COMMISSION
$6,500
2–3 weeks
Five to eight scenarios with scoring rubrics, debrief logic, and a safety review — authored for your exact practice need and delivered as YAML you own, running on hardware you control.
PRIVATE PRACTICE PILOT
$12k
6 weeks · cohort ≤ 50
A commissioned pack plus a branded deployment, a facilitator guide, and pre/post measurement design — everything one cohort needs to run the practice and prove it worked.

Both land under a single year of a 50-seat cloud contract (roughly $18k–$30k at market per-seat rates) — and unlike that contract, they ship with zero data egress. You are buying an asset you keep, not a subscription that reads your people.

For the L&D buyer whose real question is compliance, the architecture is the answer: local models, an on-device SQLite database, and the same verifiable offline receipt the skeptic asked for above. Point it at any scenario and it fails loudly if a subsystem so much as reaches for the network. One line for the procurement file: the data never leaves the building. (A regulated-sector one-pager maps this posture to HIPAA and FERPA in plain language — page plus a downloadable PDF for your privacy office.)

The candor cuts both ways. If your procurement requires a cloud dashboard of employee transcripts, we are not your vendor — on purpose. There is nothing to dashboard, by design, and we will not build the surveillance back in to win the deal.

Commission a pack or pilot Try it — $9.99 on Steam Read the source

Free to build. Fair to buy.

The engine is Apache-2.0 and the official packs are CC BY 4.0 — clone it, build it, run it for nothing, forever. The $9.99 Steam version is the same software, packaged: signed, notarized, auto-updating, Steam Deck–ready. You’re not paying to unlock the product; you’re paying for the convenience of not building it yourself — and so the work can continue. Nothing free today ever moves behind that price. The paid content is only ever more.

DEBRIEF · THE PORTFOLIO REVIEW
Said clearly: the architecture is the privacy policy. Hedged: nothing observed. Turning point: producing a verifiable receipt instead of a promise. Outcome: grudging_respect earned.

From the studio manifesto: presence over avoidance — avoidance is the root of most suffering, and you can’t out-think reality; you can only meet it. Conversation Simulator is that belief shipped as software: moments of discomfort, followed by clarity, on a loop, until the version of you that walks into the real room has already been here.

Proof: conversationsimulator.com · manifesto · docs · source (Apache-2.0) · download / Steam — Windows, macOS, Linux, Steam Deck

Other Interactive Projects

< All Projects