coco-second-brain

Second Brain, Zero Dashboard replaces the static PKM dashboard with a live, agent-generated UI that assembles itself around your thinking in real time. Instead of typing into forms or clicking through menus, you brain-dump in natural language — and a 5-stage AI pipeline responds not with text, but with a fully rendered personal dashboard built fresh for each input.

The pipeline runs as follows: a router agent classifies the domain (career, emotional, writing, research) and selects which sub-agents to invoke. Three parallel sub-agents then run concurrently — a Theme Extractor that names the core thread in your own language, an Energy Detector that reads emotional texture and energy state, and a Pattern Linker that finds non-obvious connections across prior captures in the same session. A deterministic A2UI spec generator flattens all sub-agent outputs into a flat component array. The frontend renders that spec through a typed component catalog — theme cards with weight bars, energy alerts, pattern detection badges, and action items — none of which exist until the agent decides they should.

The stack is React 18 + TypeScript + Vite on the frontend, with the Anthropic Claude API (claude-sonnet-4-6) powering all five stages. The A2UI protocol drives the generative UI layer: agents return structured JSON specs rather than prose, and the renderer interprets component type, props, and layout at runtime. Every pipeline stage is instrumented with LangSmith tracing so the full router → sub-agent → spec → render chain is observable. The result is an interface that has no fixed layout — it is entirely generated by agent output, making every brain dump produce a different, contextually appropriate dashboard.