Standalone PKM app — yjs CRDT sync stack (2026-05-06, 15:06)
Same-day follow-up (~45 min after 2026-05-06-Standalone PKM app - Cursor SDK as harness layer). Pivots the harness pick from Cursor SDK → Claude Code SDK and proposes yjs CRDTs as the sync layer (with a future “bridge for local files” — the local-first-md-vs-databases tension from challenge #3). Stack is concrete enough to prototype against.
ideas:
- use yjs CRDTs for sync layer + backend (+ bridge for local files in the future)
stack:
- yjs
- tiptap
- assistant ui (https://github.com/assistant-ui/assistant-ui) with shadcn template
- shadcn for normal components
- claude code sdk as AI harness layer + login with claude code
- https://claude.ai/share/48cde643-838d-4d41-9773-533c4928f3aa
Related
- 2026-05-06-Standalone PKM app - cons of Obsidian setup, sync, UI focus — parent memo enumerating the sync / cost / UI challenges; this stack tackles sync (#2) and cost (#1, via “login with claude code”) directly.
- 2026-05-06-Standalone PKM app - Cursor SDK as harness layer — earlier same-day pick (Cursor SDK), now superseded here by Claude Code SDK + login.
- OpenClaw for PKM — parent concept page.
- My AI native Obsidian Setup — the existing setup this would replace; the “login with claude code” idea collapses the moving parts called out as con #4 of that page.
- Open-Source Model-Agnostic AI Platform — adjacent platform framing; this stack is one concrete instantiation of that pattern.
- Agent-Native Software Architecture — the architectural pattern; Claude Code SDK as harness is the orchestrator layer.
- Kumo — inverse stack at the storage layer: pure cloud filesystem with MCP + REST + per-agent path scopes, no AI built in; usable as the backend a yjs+tiptap PKM client would persist into (replacing the “yjs as backend” line item with explicit blob storage + scoped agents).
- 2026-05-10-Inbox-style PKM app - input funnel and iOS sync challenge — same project track, 4 days later: doesn’t pick a harness but flags the “roll your own sync engine” worry this stack proposal directly addresses.
- 2026-05-19-Distill app review - AI-native PKM with local-only limits — 13 days later: Distill ships E2E-encrypted server sync with a vault password — a different point in the same sync-design space, no CRDT layer surfaced. Useful contrast to this CRDT proposal.