A daily training app, paired with an autonomy score that runs across the AI tools you already use. Practice thinking for yourself. Ten minutes a day.
The word that used to come easily doesn't.
The idea you would have had on your own. You asked an AI instead.
The problem you would have worked through. You pasted it into a chat window.
Four paradigms, paired and progressively loaded. Sessions adapt to your performance, not your streak.
Nodep runs locally inside your browser. We see the prompts you send and the responses you receive. Never the rest of your tabs, never your typing outside the chat box.
When you send a prompt, Nodep classifies it on-device against your reach-for patterns. Reflex prompts trigger a one-tap pause. Unfamiliar problems pass through clean.
Four signals. That's it. Prompt content stays on-device for classification; only the derived signals leave it.
Classification runs on a 38MB local model. The only thing we sync is the four signals above, hashed per-tool. You can export or wipe everything from settings. One tap, no ticket.
Your autonomy score reconciles signals across Claude, ChatGPT, Cursor and the rest, so the same reflex doesn't get rewarded twice for switching tabs.
Updated daily. Measures how often you reach for AI before thinking, how often you show your work, and how that ratio shifts over weeks, not days.
Most people who use it are surprised by what they see.
Each paradigm targets a separable executive function. The training app pairs them deliberately. N-back with Go/No-Go, Stroop with task-switching, to load working memory under interference and inhibitory demand simultaneously, which is what real cognitive offload looks like in the wild.
Difficulty is set per-session by an adaptive staircase calibrated to your previous five sessions, the same procedure the source studies used.
Why these four. The paradigms above cover the executive-function profile most affected by chronic cognitive offloading: working-memory load, attentional selection, response inhibition, and set-shifting.
What we measure. Task-internal accuracy and reaction-time variance feed your training trajectory. Cross-tool autonomy signals feed your score. The two are reported separately and never blended.