Collective Vision

Introduction

An AI-native, embeddable user-feedback platform that runs on Cloudflare's edge.

Collective Vision is an AI-native, embeddable user-feedback platform that runs cheaply on Cloudflare's edge — UserVoice-style boards (submit / vote / comment) with deduplication, auto-tagging, sentiment scoring, and an MCP server so AI agents can read and write feedback alongside humans.

Drop a single <script> tag onto any website — the way Disqus is embedded for comments — and start collecting feedback. Workspaces and boards auto-provision on the first hit, so there is no signup or API call required to go live.

Try it now

A working dev instance runs on Cloudflare Workers (free plan):

Embed the demo widget on any page:

<script
  src="https://feedback-dev.jfcreations.com/widget.js"
  data-workspace="demo-workspace"
  data-board="main"
></script>

What you get

  • Embeddable widget — submit ideas, upvote, comment; anonymous-by-default (the widget mints an anon_<random> id into localStorage).
  • Multi-tenant APIworkspaces → boards → feedback_items, isolated by workspace_id, with auto-provisioning on first use.
  • Moderation pipelinepending / approved / rejected plus is_hidden, with per-source defaults (widget auto-approves; MCP / API / import require review).
  • Vote deduplication — one vote per user per item, enforced in the schema.
  • AI processing — embedding + semantic deduplication run free and inline on Workers AI; LLM classification (type / sentiment / urgency / priority) is pluggable.
  • MCP server — agents query, search, submit, vote, and comment over JSON-RPC 2.0, alongside humans.

How AI classification works

Embedding and semantic dedup always run free, inline, on Workers AI. The LLM classification step can come from any of three sources:

  1. claude bridge — key-free; runs the local claude CLI on your machine. See Claude Bridge.
  2. OpenAI — set the OPENAI_API_KEY secret; runs inline on submit.
  3. Anthropic — set the CLAUDE_API_KEY secret; runs inline on submit.

Without any of these, classification is skipped by design (fields stay null, no error); embedding and dedup still run.

Architecture

Browser widget ─┐
External APIs ──┤
MCP clients   ──┼──▶  Cloudflare Worker  ──▶  D1 (SQLite)
Imports       ──┘         │                       │
                          ├──▶ Vectorize (embeddings, AI search)
                          ├──▶ Workers AI (inference)
                          ├──▶ Queue: ai-processing (async)
                          └──▶ Durable Object: PlanEditorSession

Interactive architecture map

A full interactive architecture map — routes, data model, MCP tools, and the AI pipeline across 7 tabs — ships with this site at /architecture-overview.html. Open it in a new tab.

LayerTech
RuntimeCloudflare Workers (TypeScript, ESM)
DatabaseCloudflare D1 (SQLite)
Vector searchCloudflare Vectorize (768-dim, cosine)
InferenceWorkers AI
Async workCloudflare Queues
Real-time co-editDurable Objects (PlanEditorSession)
Admin UIVite + React + Tailwind

Where to next

On this page