| name | ai-agent-projects |
| description | AI agent project scaffolding templates for vibe coders. Knows how to scaffold voice agents (Telnyx + ElevenLabs), chat agents (OpenRouter + Supabase), multi-agent systems, RAG pipelines, and MCP servers. Use when /cks:kickstart detects an AI agent project type, or when user says "I want to build an AI agent."
|
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash |
| model | sonnet |
AI Agent Projects for Vibe Coders
Scaffold AI agent projects from idea to deployable product. These are opinionated
templates matching the CKS stack defaults: Next.js + Supabase + OpenRouter +
n8n + Vercel + Railway.
Agent Project Types CKS Can Scaffold
| Type | Description | Default Stack |
|---|
| voice-agent | Phone-based AI voice agent (inbound/outbound) | Telnyx + ElevenLabs + Deepgram + Supabase |
| chat-agent | Web chat bot with memory and tools | Next.js + Supabase + OpenRouter |
| multi-agent | Multiple AI agents coordinating on tasks | Supabase + n8n + OpenRouter |
| rag-pipeline | Retrieval-augmented generation over documents | Supabase/pgvector + OpenRouter + Next.js |
| mcp-server | Model Context Protocol server (extends AI tools) | TypeScript + Fastify or Express |
| n8n-workflow | n8n-hosted agent with MCP tools | n8n + OpenRouter |
Pattern: Voice Agent Stack
Architecture
Inbound call
↓
Telnyx number → SIP trunk
↓
ElevenLabs Conversational AI (agent_7901ksp5mtc1fpkvqmeedwqmde50)
— STT: Deepgram (via ElevenLabs)
— LLM: OpenRouter (claude/gemini)
— TTS: ElevenLabs (voice EkK5I93UQWFDigLMpZcX)
↓
n8n webhook → Supabase CRM / Brain MCP
What to Scaffold
voice-agent/
├── .voice/
│ ├── config.json ← Telnyx number, agent ID, ElevenLabs config
│ ├── system-prompt.txt ← AI agent behavior prompt
│ └── knowledge-base/ ← KB documents for ElevenLabs
│ ├── faq.md
│ └── procedures.md
├── n8n/
│ └── crm-integration.json ← Webhook workflow templates
├── supabase/
│ ├── schema.sql ← Calls, transcripts, contacts tables
│ └── rls-policies.sql ← Row-level security
├── AGENTS.md ← Hermes standing orders
├── CLAUDE.md ← Claude Code standing orders
└── README.md
KB Content Rule
AI agent knowledge bases MUST use actual site content (blog/guides/copy),
never manual summaries. Point to source URLs, not generated text.
System Prompt Template (Voice)
You are a {agent-name}, a voice AI assistant for {company/project}.
Rules:
- Speak in complete sentences. Never use markdown.
- Maximum {N} sentences per response.
- Confirm before any destructive action.
- If unsure, ask ONE clarifying question.
- Your knowledge is in the KB. Don't hallucinate.
- End every successful interaction by logging to CRM.
Pattern: Chat Agent Stack
Architecture
User message
↓
Next.js chat UI (/chat)
↓
Supabase (auth, session history, message storage)
↓
OpenRouter (model routing per task type)
↓
Tool calls (n8n webhooks, Supabase queries, MCP servers)
↓
Streaming response back to UI
What to Scaffold
chat-agent/
├── src/
│ ├── app/
│ │ ├── page.tsx ← Landing + chat trigger
│ │ ├── chat/
│ │ │ ├── page.tsx ← Chat UI
│ │ │ └── api/
│ │ │ └── route.ts ← Streaming chat endpoint
│ │ └── layout.tsx
│ ├── components/
│ │ ├── ChatMessage.tsx
│ │ ├── ChatInput.tsx
│ │ └── ToolCallCard.tsx
│ ├── lib/
│ │ ├── llm.ts ← OpenRouter client
│ │ ├── tools.ts ← Tool definitions
│ │ └── supabase.ts ← DB client
│ └── types/
│ └── chat.ts
├── supabase/
│ ├── schema.sql
│ └── rls-policies.sql
├── AGENTS.md
├── CLAUDE.md
└── README.md
Pattern: Multi-Agent Stack
Architecture
User request
↓
Orchestrator agent (classifies intent)
↓
Specialist agents (parallel or sequential):
├── Researcher agent → web search + synthesize
├── Writer agent → generate content
├── Reviewer agent → validate quality
└── Summary agent → combine results
↓
n8n (orchestration + tool execution)
↓
Response to user
What to Scaffold
multi-agent/
├── src/
│ ├── orchestrator.ts ← Intent classifier + agent router
│ ├── agents/
│ │ ├── researcher.ts ← Web search + fact synthesis
│ │ ├── writer.ts ← Content generation
│ │ ├── reviewer.ts ← Quality validation
│ │ └── summarizer.ts ← Result combination
│ ├── tool-registry.ts ← Available tools per agent
│ └── types.ts
├── supabase/
│ ├── schema.sql ← Agent state, logs, results
│ └── rls-policies.sql
├── AGENTS.md
├── CLAUDE.md
└── README.md
Pattern: RAG Pipeline
Stack Selection
- Vector DB: Supabase/pgvector (free, auth integrated)
- Embeddings: OpenAI text-embedding-3-small via OpenRouter
- Chunking: LangChain or custom recursive splitter
- Retrieval: Hybrid (vector + keyword via pgvector + tsvector)
- Generation: Claude Sonnet or Gemini via OpenRouter
What to Scaffold
rag-pipeline/
├── src/
│ ├── ingest/
│ │ ├── loader.ts ← Document loader (PDF, MD, TXT, URL)
│ │ ├── chunker.ts ← Recursive text splitter
│ │ └── embedder.ts ← Vector embedding + upsert
│ ├── retrieve/
│ │ ├── vector-search.ts ← pgvector similarity search
│ │ └── hybrid-search.ts ← Vector + keyword fusion
│ ├── generate/
│ │ ├── query-rewriter.ts ← Rewrite user query for retrieval
│ │ └── answer.ts ← LLM answer with context
│ └── app/
│ └── route.ts ← Search + answer API endpoint
├── supabase/
│ └── migrations/
│ └── 001_pgvector.sql ← Enable pgvector + index
├── AGENTS.md
├── CLAUDE.md
└── README.md
Pattern: MCP Server (Model Context Protocol)
Architecture
AI agent (Claude Code / Hermes)
↓
MCP Client (stdio or HTTP transport)
↓
MCP Server (this project)
↓
Tools you define (file ops, API calls, DB queries)
↓
External services
What to Scaffold
mcp-server/
├── src/
│ ├── index.ts ← Server entry + transport setup
│ ├── tools/
│ │ ├── tool-definitions.ts ← Tool schema + handler map
│ │ ├── search-tool.ts ← Example: web search
│ │ └── db-tool.ts ← Example: Supabase query
│ ├── handlers/
│ │ └── resource-handler.ts ← Resource handler (optional)
│ └── types.ts
├── package.json ← @modelcontextprotocol/sdk
├── tsconfig.json
├── AGENTS.md
├── CLAUDE.md
└── README.md
How to Use in Kickstart
When the user types /cks:kickstart "voice agent for restaurant booking":
- Detect project type from the description (voice-agent keywords → template)
- Check
.kickstart/context.md for confirming signals
- Auto-select the matching template stack (overrideable via AskUserQuestion)
- After brand/design phases, scaffold from the template in Handoff phase
- Generate both AGENTS.md (Hermes) and CLAUDE.md (Claude Code)
Stack Selection Overrides for AI Projects
The stack-selection workflow should add these options when AI agent project
is detected:
| Category | Override Options |
|---|
| AI Gateway | OpenRouter (default), Direct Anthropic, Direct OpenAI |
| Voice Platform | Telnyx (default), Twilio, Just WebRTC (no phone) |
| TTS | ElevenLabs (default), OpenAI TTS, Edge TTS |
| STT | Deepgram (default), Whisper, OpenAI STT |
| Orchestration | n8n (default), Temporal, Direct code |
| KB Storage | Supabase (default), Pinecone, Weaviate, files |
Verification
