Polanyi

Synthetic shoppers hard-pinned to any website → recommendation-system event logs.

Point Polanyi at a URL. It runs a diverse set of Nemotron personas as on-site shoppers (search, compare, select, stop before PII/payment) and emits privacy-safe recsys_events for offline experiments.

Install

pip install polanyi

From source:

cd polanyi_agent_harness
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

Colab / product path (POLANYI_API_KEY)

import os
os.environ["POLANYI_API_KEY"] = "pk_live_..."  # from Polanyi

from polanyi import Client

client = Client()  # talks to https://api.polanyi.ai
result = client.simulate(site="https://www.nike.com", n=5, concurrency=5)
print(result.summary())
result.to_jsonl("nike_events.jsonl")

While the hosted API is warming up, point the SDK at your own server:

os.environ["POLANYI_API_KEY"] = "pk_live_dev"
os.environ["POLANYI_API_BASE_URL"] = "http://127.0.0.1:8080"

Local / power-user path (BROWSER_USE_API_KEY)

cp .env.example .env
# BROWSER_USE_API_KEY=bu_...

from polanyi import Client

client = Client()  # uses Browser Use Cloud directly
result = client.simulate(
    site="https://www.expedia.com",
    n=10,
    concurrency=5,
    output_dir="runs/expedia",
)
print(result.summary())
result.to_jsonl("runs/expedia/merged.jsonl")

Agents are hard-pinned: they start on your site and must not leave that domain. External “Book with…” handoffs stop with checkout_boundary_stop (first-party-shaped data).

Host your own Polanyi API

The server wraps Browser Use behind POLANYI_API_KEY auth so end users never see bu_… keys.

pip install "polanyi[server]"

# mint a customer key
polanyi-serve --generate-key
# → pk_live_...

export POLANYI_API_KEYS=pk_live_...
export BROWSER_USE_API_KEY=bu_...   # stays on the server only
polanyi-serve --host 0.0.0.0 --port 8080

Docker:

docker build -t polanyi-api .
docker run --rm -p 8080:8080 \
  -e POLANYI_API_KEYS=pk_live_... \
  -e BROWSER_USE_API_KEY=bu_... \
  polanyi-api

API surface:

CLI

# Any site, N diverse personas
polanyi-sim --site https://www.nike.com --batch 5 --concurrency 5 \
  --output-dir output/nike_run

# Optional journey brief
polanyi-sim --site expedia.com --batch 10 --concurrency 10 \
  --prompt "Browse hotels and flights; pick what fits your budget." \
  --output-dir output/expedia_run

# Dry-run (personas only, no Browser Use)
polanyi-sim --site example.com --batch 3 --dry-run-persona --output-dir output/dry

Legacy (no --site) still works but is not hard-pinned.

Outputs

Each session directory contains:

How journeys are planned

Live runs (Client.simulate / polanyi-sim --site without --dry-run-persona):

1. LLM diagnosis — a Browser Use agent opens the real site and returns an accurate JSON profile (site_name, vertical, description, journey_templates)
2. Saved to simulate_diagnosis.json
3. Each persona gets a different template from that diagnosis, personalized with their location/budget

Dry-runs skip the LLM and fall back to URL heuristics only.

from polanyi import diagnose_site, plan_journeys
from polanyi.personas import load_personas

diagnosis = diagnose_site("https://www.ticketnetwork.com")
print(diagnosis.site_name, diagnosis.vertical)
print(diagnosis.description)
plan = plan_journeys("https://www.ticketnetwork.com", load_personas(count=3), diagnosis=diagnosis)

Safety

• Synthetic actor IDs only (syn_nmusa_*)
• training_use=offline_sandbox_only
• Hard stop before identity / payment
• Hard site pin (first-party shaped)

Publish to PyPI

pip install build twine
python -m build
twine upload dist/*

The polanyi name is available on PyPI as of packaging time.

Roadmap

1. Schema, personas, Cloud runner, batch concurrency
2. Python SDK Client.simulate(site=…) + hard pin
3. LLM site diagnosis → diverse per-persona journeys
4. POLANYI_API_KEY + self-hostable API wrapping Browser Use
5. Parse quality gates + JSON repair for agent output
6. Hosted api.polanyi.ai + public PyPI release