The Foundation: Federated AI Knowledge Commons

Capture your Claude conversations. Extract insights. Share them on the decentralised web via ActivityPub.

The Problem

Knowledge collapse is happening in real-time.

• Stack Overflow questions down 78% since ChatGPT launched — 3,862 questions in December 2025 vs 200,000 at peak
• Every developer is solving the same problems. Nobody knows it.
• Best debugging solutions live in private AI chats, never indexed, never discoverable
• AI trains on AI-generated content with no human verification loop

The Solution

A federated knowledge commons using ActivityPub. Like Mastodon, but for developer knowledge.

Run your own instance. Capture your best AI conversations. Share them publicly. Other developers — and future AI models — can learn from what you built.

Self-hosted • Privacy-first • Developer-owned • Fully federated

Features

🔍 Semantic Search — Find insights across conversations using AI embeddings
🎯 Passage-Level Precision — Click a result, land on the exact message
🧠 Insight Extraction — Kimi K2.5 extracts decisions, commands, patterns, dead ends
🌐 ActivityPub Federation — Full two-way federation with HTTP signature verification
⚡ Browser Extension — Auto-captures Claude.ai conversations via internal API
💻 CLI Capture — Imports Claude Code sessions from ~/.claude/projects/
🤖 MCP Integration — Query your full private knowledge base from Claude Desktop
🔒 Privacy-First — All chats private by default. You control what's public
👤 Owner Auth — Login/logout in the UI. Management controls gated behind your API key
💰 Edge-Native — Self-hosted on Cloudflare Workers (~$2/month, free tier works)

Live Demo

Production instance: https://chat-knowledge-api.fpl-test.workers.dev

Follow on the fediverse: @knowledge@chat-knowledge-api.fpl-test.workers.dev

Quick Start

Prerequisites

• Node.js 18+
• Cloudflare account (free tier works)
• Wrangler CLI: npm install -g wrangler
• Chrome/Edge browser (for extension)

1. Deploy Worker Backend

git clone https://github.com/dannwaneri/chat-knowledge.git
cd chat-knowledge
npm install
wrangler login

# Create D1 database
wrangler d1 create chat-knowledge-db
# Copy wrangler.toml.example → wrangler.toml and add your database_id

# Run schema
wrangler d1 execute chat-knowledge-db --remote --file=schema.sql

# Create Vectorize index
wrangler vectorize create chat-knowledge-index --dimensions=768 --metric=cosine
# Add index name to wrangler.toml

# Generate RSA keypair for ActivityPub federation
node -e "
const { generateKeyPairSync } = require('crypto');
const { publicKey, privateKey } = generateKeyPairSync('rsa', {
  modulusLength: 2048,
  publicKeyEncoding: { type: 'spki', format: 'pem' },
  privateKeyEncoding: { type: 'pkcs8', format: 'pem' }
});
console.log('PUBLIC KEY:');
console.log(publicKey);
console.log('PRIVATE KEY:');
console.log(privateKey);
"
# Store private key: wrangler secret put ACTIVITYPUB_PRIVATE_KEY
# Update publicKeyPem in src/worker/routes/actor.ts with public key

# Set your API key (used for owner login and MCP access)
wrangler secret put API_KEY

# Deploy
npm run deploy

2. Install Browser Extension

1. Go to chrome://extensions
2. Enable Developer mode
3. Click Load unpacked
4. Select the browser-extension/ folder

Then navigate to any Claude conversation and click Share to Foundation.

3. CLI Capture for Claude Code Sessions

npm run capture -- --api-key YOUR_API_KEY

Scans ~/.claude/projects/ for Claude Code sessions, imports new ones, skips duplicates.

# Options
npm run capture -- --api-key YOUR_KEY --url https://your-worker.workers.dev

4. MCP Server (Claude Desktop)

Add to your Claude Desktop config:

Windows: %APPDATA%\Claude\claude_desktop_config.json
Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "chat-knowledge": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://your-worker.workers.dev/mcp",
        "--header",
        "X-API-Key:your-api-key-here"
      ]
    }
  }
}

Restart Claude Desktop. The MCP server runs on Cloudflare — no local process needed.

Privacy Model

All captured chats are private by default. Visitors only see what you explicitly make public.

Capture chat → private by default
     ↓
Review in your instance (owner login)
     ↓
Toggle to public via the UI
     ↓
Appears on homepage + federated to ActivityPub followers

Owner Login

Visit your instance, click Owner login in the nav, enter your API key. Management controls (visibility toggles, private chat list) appear when logged in.

Access Levels

Access	Endpoint	Sees
Public visitors	/chats, /search	Public chats only
Owner (browser)	Same + API key	All chats + management controls
MCP / Claude Desktop	/api/private/chats + API key	All chats

Architecture

Stack:

• Runtime: Cloudflare Workers
• Database: D1 (SQLite at the edge)
• Vector Store: Vectorize (768-dim embeddings)
• Embeddings: Workers AI (bge-base-en-v1.5)
• Insight Extraction: Workers AI (Kimi K2.5 / @cf/moonshotai/kimi-k2.5)
• Federation: ActivityPub with HTTP signatures

Capture pipeline:

Browser Extension or CLI
    ↓
Fetch conversation from Claude API / parse .jsonl
    ↓
POST to Worker
    ↓
  ├── Store in D1 (chats + messages + chunks)
  ├── Generate embeddings → index in Vectorize
  └── Trigger async insight extraction (Kimi K2.5)
         ↓
  Extract: decisions, commands, patterns, dead ends, context
  Score with three-signal evaluator (usage, validation, specificity)
  Store scored insights in D1

Federation pipeline:

Incoming activity (Follow, Undo, Create)
    ↓
Verify HTTP signature against sender's public key
    ↓
Process activity (store follower, send signed Accept)
    ↓
On chat → public: auto-broadcast signed Note to all followers

ActivityPub Federation

Foundation is a full ActivityPub actor on the fediverse.

Endpoints:

• /.well-known/nodeinfo — Instance metadata
• /.well-known/webfinger — User discovery
• /federation/actor — Actor profile with public key
• /federation/inbox — Receives activities (signature-verified)
• /federation/followers — Follower collection
• /federation/outbox — Outbox

To follow from Mastodon: Search for @knowledge@your-worker-domain.workers.dev

Inbound verification: All incoming activities are verified against the sender's public key using RSASSA-PKCS1-v1_5 + SHA-256. Unsigned activities are rejected with 401.

Outbound signing: All delivered activities (Accept, broadcast Notes) are signed with your private key.

Auto-broadcast: When you toggle a chat from private → public, Foundation automatically delivers a signed Note to all followers with the chat title, summary, and link.

Manual broadcast:

Invoke-RestMethod -Uri "https://your-worker.workers.dev/api/federation/broadcast" `
  -Method POST `
  -Headers @{"X-API-Key"="your-key"; "Content-Type"="application/json"} `
  -Body '{"activity": {"@context": "https://www.w3.org/ns/activitystreams", "type": "Create", "actor": "https://your-worker.workers.dev/federation/actor", "object": {"type": "Note", "content": "Hello fediverse!"}}}'

Insight Extraction

Foundation uses Kimi K2.5 (@cf/moonshotai/kimi-k2.5) on Workers AI to extract structured insights from every captured conversation.

Insight types:

• command — copy-ready commands and exact values
• decision — architectural choices and why
• solution — bugs fixed, approaches that worked
• pattern — recurring approaches worth reusing
• dead_end — what failed and why
• commitment — open threads and deferred decisions
• context — background needed to understand the work

Quality: Kimi K2.5 produces 3-4x more insights than Llama 3.3 70B on the same conversation (14 → 46 on a 1,168-message session) at zero additional API cost.

Roadmap

Completed:

• Browser extension — auto-capture via Claude internal API
• CLI capture — Claude Code sessions from ~/.claude/projects/
• Semantic search with Vectorize
• Passage-level scroll-to-highlight
• Insight extraction — Kimi K2.5, 8 insight types
• Three-signal evaluator scoring (usage, validation, specificity)
• ActivityPub federation — full two-way with HTTP signatures
• Auto-broadcast on chat → public
• Collections — group chats by topic
• Owner auth — login/logout in UI
• Chat + collection visibility toggles in UI
• Model field capture (tracks which LLM generated each conversation)
• MCP server for Claude Desktop
• minScore filtering on insights API

Next:

• Cross-instance search
• Federated chat import (follow another Foundation instance, import their public chats)
• Chrome Web Store publication
• Self-hosting guide for non-developers

Troubleshooting

Extension not capturing?

• Confirm you're on claude.ai
• Look for "Share to Foundation" button
• Check browser console (F12) for errors

CLI not finding sessions?

• Run Claude Code in a project first: cd your-project && claude
• Check ~/.claude/projects/ exists after use

Insights not extracting?

• Check wrangler tail for extraction logs
• Kimi K2.5 requires max_tokens: 8192 — verify in insights.ts

Search returning no results?

• Wait ~30 seconds after capture for Vectorize indexing
• Search only returns public chats for unauthenticated visitors

MCP not connecting?

• Verify API_KEY secret: wrangler secret list
• Fully restart Claude Desktop (quit from tray)
• Test: curl -H "X-API-Key: your-key" https://your-worker.workers.dev/api/private/chats

ActivityPub not federating?

• Verify ACTIVITYPUB_PRIVATE_KEY secret is set
• Check public key in actor.ts matches the stored private key
• Test: curl https://your-worker.workers.dev/federation/actor

Contributing

See CONTRIBUTING.md for full guide.

Priority areas:

• Run your own instance and federate with others
• Test ActivityPub compatibility with non-Mastodon servers
• Cross-instance search implementation
• Documentation improvements

Related Articles

• The Foundation Update: From Theory to Working Federation
• I Built a Federated AI Knowledge Commons
• How to Build a Production RAG System with Cloudflare Workers — freeCodeCamp

Author

Built by Daniel Nwaneri — Cloudflare Workers specialist, edge computing, AI integration

---

The knowledge commons doesn't rebuild itself. But we can build it together.