noticed-claw

a multi-tenant agent workshop built on noticed's production architecture.

what is this

a self-contained Next.js multi-tenant agent architecture where each user gets their own persona, memory, workspace files, scheduled jobs, and conversation history.

see WORKSHOP.md for the full guide.

quick start

# 1. install
npm install

# 2. start local supabase (requires docker)
supabase start

# 3. run database migrations
for f in src/supabase/migrations/*.sql; do
  psql postgresql://postgres:postgres@127.0.0.1:54322/postgres -f "$f"
done

# 4. seed 100 developer profiles
npm run seed

# 5. copy env and fill in keys
cp .env.example .env
# fill in supabase keys (from `supabase status`) and OPENAI_API_KEY

# 6. start dev server
npm run dev

open http://localhost:3012 — sign up, then chat.

what's inside

src/
├── app/
│   ├── api/
│   │   ├── chat/            streaming chat endpoint (AI SDK useChat)
│   │   ├── sessions/        session CRUD + message history
│   │   ├── webhook/         platform adapters (telegram, slack)
│   │   ├── agent-cron/      heartbeat + cron tick
│   │   └── eval/            eval runner endpoint
│   ├── dashboard/
│   │   ├── chat/            webchat with session selector
│   │   ├── developers/      browse 100 seeded profiles
│   │   ├── evals/           eval runs + per-scenario detail
│   │   └── tenants/[id]/    sessions, messages, memories, missions,
│   │                        workspace, cron, tools
│   └── login/               email/password auth
├── lib/agent/
│   ├── agent-turn.ts        turn orchestration (non-streaming)
│   ├── stream-bridge.ts     streaming with silent-token buffering
│   ├── agent-router.ts      tenant + session resolution
│   ├── prompt-builder.ts    system prompt assembly (9 sections)
│   ├── llm-runner.ts        tool-call loop via AI SDK generateText
│   ├── memory-extract.ts    LLM-based fact extraction per turn
│   ├── memory-flush.ts      pre-compaction memory preservation
│   ├── memory-manager.ts    cosine dedup, pgvector recall
│   ├── compaction.ts        token threshold, chunked summarization
│   ├── brand-voice.ts       immutable style rules
│   ├── persona-catalog.ts   ari / donna / ted
│   ├── workspace-files.ts   7 per-tenant .md files
│   ├── mission-engine.ts    onboarding → audience → outreach
│   ├── session-awareness.ts cross-session context
│   ├── heartbeat.ts         timezone-aware proactive check-ins
│   ├── cron.ts              at / every / cron scheduling
│   ├── virtual-fs.ts        read-only fs over seeded data
│   ├── thread-queue.ts      webhook dedup via advisory locks
│   └── tools/               9 built-in tools + code mode
├── eval/
│   ├── cli.ts               npm run eval (+ CSV output)
│   ├── runner.ts            create tenant, play turns, judge, cleanup
│   ├── judge.ts             6-dimension LLM scoring
│   └── scenarios/           10 YAML eval scenarios
└── supabase/
    ├── client.ts            browser + service role clients
    ├── auth-client.ts       cookie-based auth (server only)
    ├── seed.ts              100 developer profiles with embeddings
    └── migrations/          14 SQL files (13 tables + RLS)

commands

command	what it does
npm run dev	start dev server on :3012
npm run build	production build
npm run test	run 197 vitest tests
npm run lint	eslint with 4 architecture rules
npm run seed	seed 100 developer profiles
npm run eval	run 10 eval scenarios, output CSV

eval system

npm run eval

scores scenarios across 6 dimensions (0-10): coherence, persona adherence, tool usage, brand voice compliance, task completion, memory quality.

outputs CSV to eval-results/ for comparison across runs.

tech stack

• Next.js 15 (App Router, Turbopack)
• Supabase (Postgres, pgvector, RLS, auth)
• AI SDK (ai + @ai-sdk/openai) — provider-agnostic LLM + embeddings
• Vitest — unit + integration tests
• Tailwind CSS — dashboard styling
• ESLint — 4 custom architecture enforcement rules