Problem
Groups doing serious work lose important context across browser tabs, voice notes, photos, files, links, and private conversations. Most collaboration tools either centralize raw context on hosted platforms or jump straight from capture to AI output without enough human review. Coop starts from a different premise: the shared memory of a group should be earned through explicit review, not assumed by default.
Architecture
Coop is a browser-first, local-first system built as a Bun monorepo. The MV3 extension is the primary work surface for capture and review, the receiver PWA handles mobile and secondary device intake, the shared package owns schemas and domain flows, and a small Hono API supports signaling and optional Yjs document sync. Raw capture stays local first; reviewed drafts, artifacts, board views, and proof material only enter shared coop state after a member chooses to publish them.
Development
The work has been about turning agentic software into a trustworthy group workflow. I shaped the capture, review, and publish loop; hardened receiver pairing and private intake; worked through local AI refinement across WebGPU, WASM, and heuristic tiers; and kept the release path honest by separating deterministic test proof from real browser and extension proof. The product is staged for browser-first use while onchain, archive, and privacy rails remain gated behind explicit modes.
Learnings
Coop has sharpened how I think about local-first AI. The hard part is not only running models near the user, it is deciding what should become shared memory at all. A useful agent has to preserve context, expose uncertainty, and make review feel lighter without removing human judgment. That has made Coop a practical bridge between my infrastructure work and the daily coordination problems teams actually feel.