the course & your data

Privacy & Transparency

This page explains, completely and in plain language, what happens to everything you type into the course's interactive simulators. Nothing on this page is boilerplate — it describes how the software actually works.

The short version

Your work stays in your browser. Nothing you build in the exercises is saved on our server.
No accounts, no analytics, no tracking. The site doesn't know who you are.
Prompts are never logged by the course. They pass through our relay to the model provider and are gone.
No model trains on what you type. The routes we use are no-training API routes (details below).
Still: use made-up data. Don't put personal or proprietary information into prompts — synthesize test data instead.

Where your work lives

Every simulator — Agent Forge, Agent Lab, the Tool Playground, the Memory Lab, the Relay, the Skill Builder, all of them — keeps its state in your browser's localStorage: your tools, saved agents, conversations, uploaded-file text, and settings. This storage belongs to your browser profile on your machine. It is never uploaded, synced, or backed up by us, and we cannot see it.

Two practical consequences. First, your work follows the browser, not you — a different machine or browser starts empty. Second, a private/incognito window is a complete privacy tool here: do the exercises in one, close it, and the browser discards everything the site stored.

You can also clear everything explicitly, right now, from this page:

Checking what this browser has stored…

Exactly what this clears

af:* — Agent Forge — tools, saved agents, runs, filesystem
al:* — Agent Lab — saved agents and conversations
ar:* — The Archive — your documentation trees and run history
cs:* — Conversation Simulator — your script
ds:* — Design Studio — your design and analysis rounds
ml:* — Memory Lab — chat, memory files, prompts
pt:* — Pattern Trainer — patterns and saved sets
rl:* — The Relay — runs, methodologies, archive
sb:* — Skill Builder — skills in progress (and the shared model setting)
tp:* — Tool Playground — tools and transcripts
wl:* — Window Lab — runs and settings
seq:* — Home page — view-mode preference
ob:* — Offering builder — view-mode preference
offering:* — Offering pages — view-mode preference
talk:* — Talk view — presentation run state

All of it lives in your browser's localStorage for this site only. Nothing here touches the server — the export file is built in your browser and downloads straight to your machine (with any API key stripped out first), and import reads the file locally too.

What our server does — and does not — see

Almost everything runs inside the browser tab, including the sandboxed execution of tool code in Agent Forge. The server side of this site does exactly three things:

Serves the pages — the lessons, exercises, and offerings you read.
Relays model requests. When an exercise sends a prompt in the default (instructor-paid) mode, the browser posts it to /api/llm, which attaches the instructor's private API key and forwards the request to OpenRouter. The relay does not write prompts or completions anywhere — no database, no log statement. It exists only so the key never has to be in your browser.
Proxies tool web requests. If you forge a tool that fetches a web page, browsers' cross-origin rules (CORS) would block most direct requests — so the tool's fetchText(url) goes through /api/fetch, which retrieves the page server-side and hands the text back. Again: pass-through, not storage.

Like essentially every website, the hosting platform (Vercel) keeps routine, short-lived operational logs of HTTP requests (IP address, URL, time). Those are infrastructure logs, not content — prompt bodies are not written to application logs. The course's database holds only course structure (the offering pages) and the instructor's own password-protected lecture recordings — no student work.

Where prompts go, and who can keep them

Prompts travel: your browser → our relay → OpenRouter → the model's provider (OpenAI, Anthropic, and others, depending on the model selected in the exercise's settings). What each party does:

This site: nothing retained, as described above.
OpenRouter: by default OpenRouter does not store prompt or completion content — it keeps only request metadata (timestamps, model name, token counts). Prompt logging is an opt-in account feature, and the course account does not opt in. See OpenRouter's data-retention documentation.
OpenAI (the default models): API traffic is not used to train models. OpenAI temporarily retains API inputs/outputs for up to 30 days solely for abuse monitoring, then deletes them (OpenAI's data-controls policy). Full zero-retention endpoints exist but require an enterprise agreement, so we do not claim them.
Anthropic (Claude models): the same shape — API content is not used for training under Anthropic's commercial terms, and inputs/outputs are deleted within about 30 days (Anthropic's retention policy).
Other providers (Kimi, GLM, DeepSeek in the model menu): policies vary by provider. OpenRouter labels each endpoint's data policy and, unless an account explicitly enables it, does not route to endpoints that train on inputs. If this matters for your experiment, stick to the OpenAI or Anthropic models — or ask.

Summed up honestly: nothing you type is used to train any model, and nothing is kept anywhere longer than the providers' ~30-day abuse-monitoring window — and most of it is kept nowhere at all.

Who pays, and using your own key

In the default mode, the instructor pays for every token through a course OpenRouter account. If you prefer that your prompts never even touch the course relay, every simulator's ⚙ settings panel accepts your own OpenRouter API key. In that mode your browser talks to OpenRouter directly — our server is not in the path at all. You can create a key at openrouter.ai/settings/keys (a small prepaid balance is enough for the whole seminar). The key is kept in your browser's storage, sent only to OpenRouter, and removed by the clear-button above. Keys from OpenAI or Anthropic directly are not currently supported — OpenRouter is the one integration point.

Cookies & tracking

There are no analytics, no tracking pixels, no third-party scripts, and no advertising anywhere on this site. The only cookie the site ever sets is the unlock cookie for password-protected offering pages — it proves to the server that you entered the access password, identifies nobody, and expires after 30 days. Everything else uses the browser storage described above, which never leaves your machine.

What we ask of you

Even with all of the above, treat the simulators like any external AI service: please don't put personal data (yours or anyone else's), confidential research data, or proprietary information into prompts. Experiment with synthetic or pseudonymized data instead. A practical trick: ask an LLM you already use — ChatGPT, Claude, Gemini — to generate a realistic fake dataset (“make me a CSV of 50 fictional patients with plausible Norwegian names, ages, and admission dates”), and experiment on that. Synthetic data behaves the same in every exercise and carries no risk.

If you'd rather not use the simulators at all

Wherever possible, the exercises come with equivalent guides for running the same experiment in a chat tool you already use and trust — ChatGPT, Claude, or Gemini — under your own account and that provider's terms. Some of the more mechanical labs (context windows, live tool loops) are hard to reproduce faithfully in a plain chat window, but we have tried to provide a workable alternative for each; where one exists, the lesson links to it. Using the alternatives affects nothing about participation in the seminar.

Questions

If anything here is unclear, or you want a deeper look at how a specific exercise handles data, ask — the entire courseware is readable source code, and showing exactly what it does is part of the course's point.

This page describes the system as deployed and is updated together with the code it describes.