一个给创作者和构建者的 agentic space。

Last edited just now · 18 min read

Vibecape is a local-first workspace where agents and humans co-write. The goal for the next quarter is to make the agent feel like a real collaborator: it remembers context across pages, suggests edits in the margin, and never interrupts the flow with noisy notifications.

Q3 Goals

The Q3 cycle is focused on three pillars: reliability, context, and polish. Each pillar has a set of deliverables that roll up into the public beta launch at the end of the quarter.

• Refactor the document model to support block-level CRDTs
• Add inline agent suggestions with accept / reject / modify actions
• Ship sync protocol v2 with encrypted peer-to-peer replication
• Build the extension runtime for custom agent tools
• Mobile companion app for capture on the go
• Public beta waitlist and onboarding flows

Design Notes

The interface should stay quiet. No noisy notifications, no infinite scroll, no competing sidebars fighting for attention. Just a tree on the left, a page in the middle, and a conversation on the right.

Agents should feel like co-authors, not chatbots. They write in the margin, suggest edits, and carry context across pages without being asked. When an agent is thinking, the UI should show a calm indicator. When it acts, the change should be visible immediately.

Motion and Feedback

Transitions should be fast and purposeful. The sidebar slides, the baybar resizes without reflowing the main content, and tool outputs expand smoothly. Every motion gives the user a sense of where things live.

Technical Direction

The core runtime is built on a local SQLite database with CRDT-backed conflict resolution. Documents are stored as a graph of blocks, so agents can edit a single paragraph without rewriting the whole page.

Sync happens opportunistically. When the device is online, changes replicate to trusted peers. When offline, the app keeps working with no degradation. Encryption is end-to-end, and keys never leave the device.

Extension Runtime

Extensions are sandboxed Web Workers that expose typed tools to agents. A tool can read from the document graph, call external APIs, or write changes back through the same permission model the UI uses.