Roast your app in 30 seconds

npx truman-cli roast --target https://your-app.com

Three brutal personas. One command. A report that hurts your feelings.

You describe people. Truman puts them in your app.

They tap, scroll, rage-quit, come back tomorrow, forget their password, and accidentally find the feature you almost deleted.

Then you get a report on what worked — and what made them leave.

Meet your worst users

# Gen Z intern. Mass-closes tabs if anything takes over 2 seconds.
- name: Jaden
  patience: 1
  persona: "If it doesn't load instantly, it doesn't exist."
  quirks: ["Swipes before the page finishes rendering"]

# Your mom trying to use your app. Reads every word on screen. Every. Single. Word.
- name: Linda
  patience: 4
  persona: "Reads the Terms of Service. Clicks 'Learn more' on every tooltip. Still not sure what the app does."
  quirks: ["Taps the logo expecting it to do something", "Reads every word on screen"]

# Senior engineer who treats your app like a pentest.
- name: Wei
  patience: 5
  persona: "Pastes Unicode into every input. Tries SQL injection on the search bar. Files issues in his head."
  quirks: ["Opens DevTools before anything else"]

# 89-year-old grandmother. Patient. Persistent. Will absolutely break your app.
- name: Abuela Carmen
  patience: 5
  techSavviness: 1
  persona: "Doesn't know what she tapped. Doesn't care. Somehow ended up on the admin panel."
  quirks: ["Double-taps everything", "Holds buttons like a TV remote"]

Frustration > 0.85 → they leave. Just like your real users do — except these ones tell you why.

Install

npm install @parentos/truman

Or run it directly — no install needed:

npx truman-cli

Quickstart

npx truman-cli init my-app && cd my-app

Point adapter.json at your API, then:

npx truman-cli run --once

Or try the included TodoMVC example — two terminals, two minutes.

Two commands. Four synthetic users. One uncomfortable truth about your checkout flow.

How it works

You describe a person.
The LLM decides what they'd do.
Your app finds out the hard way.

Scheduler  → who acts now?
Adapter    → what can they see?
LLM        → what would this person do?
Adapter    → do it
State      → update frustration, memory, discovered features

Frustration > 0.85 → they leave.

Features

• Persona-driven — personality, patience, quirks. The LLM does the rest.
• Frustration modeling — if your UX sucks, they leave. Just like real users.
• Stateful memory — they remember what broke yesterday.
• Realistic schedules — 7am coffee check, not a 3am bot swarm.
• Multi-provider — OpenAI, Anthropic, Ollama (free & local).
• Any HTTP API — adapter pattern, no SDK lock-in.
• Voice narration — hear your NPCs complain out loud.
• Bug export — pipe findings into any bug tracker.

Commands

| Command | What it does | |---|---| | truman roast | One command, 3 brutal personas, bug report | | truman run | Run simulation — NPCs use your app | | truman init | Scaffold adapter + families for your app | | truman validate | Check family YAML configs | | truman preview | Dry run — see what NPCs would do, no API calls | | truman report | Re-generate report from saved logs |

Full reference with all flags and examples: docs/commands.md

LLM providers

OPENAI_API_KEY=sk-...    truman run --once              # OpenAI (default)
ANTHROPIC_API_KEY=sk-... truman run --once -p anthropic # Anthropic
                         truman run --once -p ollama    # Ollama (free & local)

Programmatic API

import { SimulationEngine, HttpApiAdapter, createProvider } from '@parentos/truman';

const engine = new SimulationEngine({
  families: ['./families/smiths.yaml'],
  adapter: new HttpApiAdapter('./adapter.json'),
  llmProvider: await createProvider({ type: 'ollama', model: 'llama3.1' }),
});

engine.on((event) => {
  if (event.type === 'member:frustrated') console.log('User quit:', event.memberId);
});

await engine.runOnce();

What Truman is NOT

• Not a load tester — use k6
• Not an E2E framework — use Playwright
• Not a monitoring agent — use Datadog

Truman is the closest thing to putting real users in front of your app without actually doing it.

Built with

Built by the team behind ParentOS — where we test our own app with synthetic families every day.

Philosophy · Contributing · Code of Conduct · License