hafermilch¶
Multi-persona LLM agents that autonomously browse and critique your product's UI/UX — in their own voice.
Python 3.11+ · 100+ LLM providers via LiteLLM · Playwright · agent-browser · uv · hatchling
What is hafermilch?¶
hafermilch lets you define personas — characters with distinct backgrounds, goals, and expertise levels — and point them at any website. Each persona is backed by an LLM and a real browser. They navigate autonomously, form opinions in their own voice, and return structured scores and findings.
You define a persona once. You run it against any product by writing a plan.
How it works¶
flowchart LR
P[Persona YAML\n🧑 background · goals · LLM] --> R
L[Plan YAML\n🗺️ URL · tasks · steps] --> R
R[EvaluationRunner] --> B[Browser Agent\nPlaywright or agent-browser]
B -->|page context| LLM[LLM Provider\nOpenAI · Gemini · Ollama]
LLM -->|BrowserAction| B
R --> Rep[Report\nJSON · Markdown]Built-in personas¶
Senior Engineer
High technical literacy. Notices API inconsistencies, error handling gaps, and developer-experience friction.
Office Clerk
Non-technical. Confused by jargon, overwhelmed by dense layouts. Represents the median business user.
Startup Founder
Time-pressured. Cares about time-to-value, pricing clarity, and whether the product solves a real problem.
Example output¶
──────────────── hafermilch v0.1.2 ────────────────
Plan: saas_onboarding
Target: https://example.com
Personas: Senior Engineer, Office Clerk, Startup Founder
Results
┌──────────────────┬───────────┬──────────┐
│ Persona │ Score │ Findings │
├──────────────────┼───────────┼──────────┤
│ Senior Engineer │ 7.2 / 10 │ 18 │
│ Office Clerk │ 5.8 / 10 │ 14 │
│ Startup Founder │ 6.4 / 10 │ 16 │
└──────────────────┴───────────┴──────────┘
Reports written to reports/
Reports are written to reports/report.json, reports/report.md, and reports/report.html.