In every team, part of the know-how is written down nowhere. It lives in a few colleagues' heads. A newcomer gets stuck on a step, searches, asks on Slack, waits, and the customer waits too. The next newcomer pays the same friction again.
it catches these hidden frictions, unblocks people in seconds, and fixes each one once.
To test it, we built a real support desk: PipingTable, booking software for restaurants, running on Crisp, HubSpot, Slack and Notion, with fictional customers.
How it works:
- Catch. Screenpipe records the screen locally. Our agent spots a detour, such as a HubSpot search with zero results while a customer is waiting in Crisp. It sends only those few seconds to the model, with identifiers masked. The model decides whether this is a real friction or normal work.
- Unblock. If the team already has a solution, it appears in a Screenpipe notification. In rehearsal, the customer record opened 10 seconds after the search. If nobody on this computer knows, an anonymous, end-to-end encrypted request goes to the team. A colleague clicks Share without ever knowing who asked.
- Fix it once. At night, the agent remembers. It turns what the team shared, and what happened after each unanswered friction, into a general rule with no names. In the morning, the manager certifies it with a signature. The next person stuck on the same step gets the certified answer instantly.
When someone leaves, the team key changes: they can no longer read anything, and their rules stay with the team. A team map shows where the process gets stuck, never who. We measure the process, never the person.
Stack:
- TypeScript on Bun, with no native code and no npm dependency in the core.
- Supabase (Postgres, row-level security, security-definer RPCs only), or a local SQLite server, as a zero-knowledge store.
- WebCrypto: AES-256-GCM, HKDF, Ed25519 signatures.
- The Crisp and HubSpot APIs.
Screenpipe integration, through its local API only:
/searchgives us the screen text (OCR and accessibility) that the daemon cuts into short episodes. The night loop also uses it to reread the 15 minutes after a friction./v1/chat/completions, Screenpipe's AI gateway, runsclaude-opus-5to judge each episode (median 2.6 s) and to write the night rules. Identifiers are masked before sending.- Notifications (
/notify) with action buttons deliver help, "a colleague is stuck: Share", rule certification and the "Helped / Not this" feedback. - Live Views show "My agent, live" (each step), the team friction map, and live frictions (pending, resolved, certified).
Privacy: everything is encrypted on the computer before it leaves. The server only stores same-size blobs with no author, no team name and no exact time.
On Mac, api notification buttons need a small Screenpipe fix, which we propose upstream.