Vercel Templates: AI & Architecture

A curated pass over the Vercel template catalog, split into templates that ship an AI product and templates that hand you architecture you'd otherwise spend months building. Scan the table, then read only the entry you're reaching for.

Pick by what you're building

AI product templates

eve Personal Agent

Best for: a serious cross-platform personal or workplace assistant.

Persistent web conversations · Slack, iMessage, and Linear integrations · durable sessions · long-term memory · user approval before a memory is saved · authentication and account linking.

More architecturally ambitious than a normal chatbot: the same agent identity and memory follow the user across every interface. Its approval-before-save memory model is the same instinct as a self-maintaining LLM wiki — the human stays the editor.

Verdict: one of the catalog's most complete visions of an actual AI product.

Morphic AI Answer Engine

Best for: a Perplexity-style research or vertical-search product.

Cited search · multiple search and model providers · streamed generative UI · authentication · PostgreSQL history · file uploads · guest mode · shareable results.

Verdict: probably the best starting point for a research assistant, market-intelligence tool, shopping researcher, or specialized answer engine.

Open Agents

Best for: coding agents, or any agent that must work asynchronously.

Three distinct layers:

Web product → Durable agent workflow → Isolated sandbox

Repository cloning · shell and filesystem tools · hibernating sandboxes · persistent execution · cancellation · streaming · commits, pushes, and optional PR creation. (If it's opening PRs on your behalf, the git reference covers the recovery moves worth knowing.)

Verdict: both a compelling product starter and possibly the most valuable agent-architecture reference in the catalog. See Architectural leverage.

eve Content Agent

Best for: AI embedded into an existing business workflow.

Writers interact through Slack; it retrieves material from Notion, follows editable house-style skills, deterministically checks the result, and publishes approved content back to Notion.

It also demonstrates:

• User-scoped OAuth instead of shared credentials
• Approval buttons inside Slack
• Tool-specific skills
• Vercel Blob asset storage
• Sandbox execution
• OIDC authentication with no static infrastructure keys

Verdict: a better model for useful enterprise AI than "put a chatbot next to the dashboard."

MotherDuck Embedded Dives

Best for: AI-powered analytics products.

Users describe changes to dashboards, presentations, and interactive data apps in natural language. The generated interfaces stay real React components backed by live SQL.

Verdict: one of the most interesting examples of AI generating persistent product interfaces rather than text. The bundled auth is explicitly demo-grade — production needs real identity, authorization, rate limiting, and data isolation.

Next.js OpenAI Doc Search

Best for: customer support, internal knowledge search, documentation assistants.

The complete basic RAG pipeline:

1. Process Markdown/MDX documents
2. Generate embeddings
3. Store them in Supabase Postgres with pgvector
4. Retrieve relevant passages
5. Stream a grounded response

Verdict: useful and understandable, but older and narrower than the newer agent templates — take its ingestion architecture as reference and modernize the model/API layer. For what a grounded pipeline actually demands in practice, see the RAG transit planning note.

Durable iMessage AI Agent

Best for: AI delivered through texting instead of another app.

iMessage webhooks · Chat SDK · AI Gateway · tools · durable workflows · retries · structured observability. Generation and message delivery are separate retryable steps, so a transient delivery failure doesn't rerun the model.

Verdict: a genuinely useful architecture for appointment assistants, concierge services, field operations, and consumer agents.

Hume Empathic Voice Interface

Best for: voice-native assistants, coaching products, interviews, conversational experiences.

A focused implementation of Hume's real-time voice interface — an interface, not a complete business product.

Verdict: excellent voice foundation; pair it with durable workflows and persistent user state to make it a product.

Architectural leverage

Open Agents: the agent is not the sandbox

Agent state and orchestration live outside the isolated execution environment. That one separation buys:

• Durable execution beyond a request timeout
• Independent sandbox hibernation and restoration
• Replaceable model and sandbox providers
• Safe execution of generated code
• Reconnection to tasks already running

Applies to coding agents, browser agents, data agents — anything where the work outlives the request.

eve Personal Agent: channels as interfaces

Web
Slack ─────→ Shared identity, memory and tools
iMessage
Linear

The UI/API and the agent runtime deploy as separate Vercel services, which is what makes adding a fifth channel a config change rather than a rewrite.

Multi-Zones: use it deliberately

Microfrontends introduce real coordination costs. next-forge is the same trade at the monorepo level: right for a company, excessive for a weekend prototype.

Revised top five

Ranked on combined usefulness and architectural value:

1. Open Agents — best agent-system architecture
2. eve Personal Agent — best cross-channel agent product
3. Platforms Starter Kit — best general SaaS architecture
4. Morphic — best immediately useful AI product foundation
5. Liveblocks Starter Kit — best collaborative-product foundation

The combination worth building

Together: the bones of a genuinely substantial AI SaaS rather than another chatbot demo.

Caveats to carry forward

• MotherDuck auth is demo-grade — replace it before production.
• Doc Search predates the agent era; keep the ingestion, modernize the model layer.
• Hume EVI is a voice interface, not a product — it needs state and workflows around it.
• Multi-Zones / next-forge solve organizational problems; adopting them without the organization just adds coordination cost.
• Template catalogs move. Treat the architecture notes as the durable part and re-check the links before committing to any one starter.

See also

• LLM Wiki — the memory-and-approval loop that eve's agent memory implements as a product
• RAG transit planning — what a grounded retrieval pipeline demands once it has to cite its sources
• AI-Augmented Software Engineering — why context quality, not model quality, is the binding constraint on agents like these
• Git — the recovery commands worth knowing before an agent pushes on your behalf