Synk AI

Multiplayer canvas (Vite + React + Liveblocks) with multi-model AI blocks, plus an orchestration & memory backend (Express + Supabase) for sequential multi-agent handoffs.

Frontend (canvas)

npm install
npm run dev

Open http://localhost:5173 — keys live in .env.local (VITE_LIVEBLOCKS_PUBLIC_KEY, GEMINI_API_KEY, GROQ_API_KEY).

Orchestration & memory backend

The backend under backend/ routes work across pluggable AI agents and logs shared team memory in Postgres (Supabase). It does not replace the canvas AI Block /api/run path — that stays on Vercel for in-canvas Gemini/Groq runs.

1. Apply Supabase migrations

In the Supabase SQL editor (or CLI), run in order:

1. supabase/migrations/001_workspaces.sql
2. supabase/migrations/002_orchestration.sql

Confirm tables exist: workspaces, agents, memory_events, orchestration_runs, orchestration_steps.

2. Configure env

Copy backend/.env.example → backend/.env and fill:

• SUPABASE_URL
• SUPABASE_SERVICE_ROLE_KEY (preferred for server inserts)
• ANTHROPIC_API_KEY
• GEMINI_API_KEY (can reuse the same key as the frontend)

The server also loads the repo-root .env.local if present.

Optional: GEMINI_MODEL if your key rejects gemini-2.0-flash (e.g. set gemini-3.6-flash).

3. Install & start

cd backend
npm install
npm run dev

API: http://localhost:3001 — health check: GET /health

4. Seed the demo agent chain

cd backend
npm run seed
# or: npm run seed -- <existing-workspace-uuid>

This registers three agents:

Agent	Provider	Role
story-writer	Anthropic (Claude)	Brief → user stories
design-drafter	Gemini	Stories → wireframe description
dev-scaffolder	Anthropic (Claude)	Wireframe → React component breakdown

The script prints the workspaceId and three agent IDs.

5. Test the orchestration chain (curl)

curl -X POST http://localhost:3001/api/orchestrate \
  -H "Content-Type: application/json" \
  -d '{
    "workspaceId": "<demo-workspace-id>",
    "agentIds": ["<story-writer-id>", "<design-drafter-id>", "<dev-scaffolder-id>"],
    "triggerContent": "Users need a faster mobile checkout flow"
  }'

Expected: one brief → three sequential handoffs (Claude → Gemini → Claude). Each step’s output is logged to memory_events and returned in the JSON steps array.

Other orchestration routes

Method	Path	Purpose
POST	/api/agents	Register an agent
GET	/api/agents/:workspaceId	List agents
POST	/api/orchestrate	Run an agent chain
GET	/api/memory/:workspaceId	Team memory summary
POST	/api/generate-stories	Convenience single-agent endpoint (migrated to orchestrator.runAgent)

Layout

api/run.ts              # Vercel serverless — canvas AI Block Gemini/Groq
server/runAi.ts         # Shared canvas AI helpers
backend/                # Express orchestration + memory layer
supabase/migrations/    # Postgres schema for agents & memory
src/                    # React canvas (Liveblocks) — do not confuse with backend