AI Agent Starter Kit for Production SaaS

An AI agent starter kit is a pre-built foundation for shipping SaaS products with AI agent capabilities: tool calling, multi-provider LLM routing, streaming chat, usage tracking, and multi-tenant isolation. VibeReady ships it as a Next.js AI agent template you customize for your domain instead of building AI infrastructure from scratch.

Streaming chat built on Vercel AI SDK, multi-provider LLM routing via OpenRouter (300+ models including OpenAI, Anthropic, and Google), a tool calling system with role-based access and schema validation, multi-tenant isolation scoped by organization, and token usage tracking per user and per org.

Describe the feature to your AI coding tool and vibe-code it on top of the kit's LLM-ready architecture. The AI framework loads the right context automatically (scoped rules, module READMEs, and skills like new-feature), so generated code follows the shipped patterns: org-scoped tools, Zod-validated parameters, and approval gates where a write needs one. Quality gates run tests and type checks before anything lands, whether you're adding a domain tool, a new data source, or a whole agent workflow.

Every tool call receives the user's organizationId and userId. Tools can only query data within that organization. There's no way for one user's agent to access another organization's data — isolation is enforced at the tool level, not the prompt level. This is built into the factory pattern, not bolted on after the fact.

OpenRouter gives your agent access to 300+ models, including OpenAI, Anthropic, and Google, through a single API. You can also connect directly to OpenAI, Anthropic, or Google APIs. Switching providers is a config change (LLM_PROVIDER plus the provider's API key) with no code changes, and within OpenRouter, switching models needs no config change at all.

The AI Framework includes development tools: structured workflows, quality gates, and architectural context for building with AI coding tools. The Full Kit adds user-facing agent infrastructure: Vercel AI SDK, OpenRouter integration, tool calling system, chat UI components, plus all SaaS features (auth, billing, multi-tenancy, infrastructure). Agent features require the Full Kit.

LangChain and CrewAI are agent orchestration frameworks — they help you build agents but not SaaS products. VibeReady is a SaaS starter kit with agent infrastructure built in: authentication, billing, multi-tenancy, deployment pipeline, plus tool calling and LLM routing. You ship a product with a login page, billing, and AI features — not a Python script.

The kit's shipped MCP integration is on the development side: Context7 (always-current library docs) and Chrome DevTools (live browser debugging) come pre-configured for your AI coding tools. For the in-app agent, MCP is an extension you can build on the AI SDK foundation rather than a wired-in feature; the shipped path for connecting your own data is the knowledge base (RAG) and the tool registry.

Yes. Threads persist in a sidebar history with auto-generated titles, so users can resume any past conversation and the tool calls replay in context. Beyond per-thread history, there's cross-session AI memory scoped per organization and user — the agent recalls facts and preferences across separate chats, with automatic context compaction so long histories stay within the model's window. Persistence is built into the Full Kit, not a plugin you bolt on.

Yes. Write-actions — invite a user, change a role, remove a member — run through a Human-in-the-Loop approval step. The agent proposes the action and an owner or admin sees an inline Approve or Deny prompt before anything changes, with an AI-provenance audit trail recording what the agent did. Tools are guarded by Zod schemas and role checks, so a member's agent can't trigger an owner-only action. Only 6% of companies fully trust AI agents to run core processes (HBR Analytic Services, 2025), which is exactly why every team write-action stays gated behind a human.

The Full Kit ships an AI Usage & Cost Dashboard (owner/admin) that records real per-request cost from OpenRouter, token counts, and a breakdown by model and tool — exportable to CSV. Usage is tracked per user and per organization, so you can see exactly what each customer costs, wire it into metered billing, or set your own caps. On OpenRouter, cost is recorded per request at the source, so the numbers reflect actual spend; other providers use sourced list prices, and anything unpriced is disclosed as such rather than guessed.

Yes. The Full Kit's assistant works as an in-app AI copilot for your SaaS: it streams answers over your product's real data (analytics, billing, team, notifications) through role-checked tools, and proposes write-actions that owners or admins approve inline. You rebrand the chat UI and add domain tools instead of building copilot infrastructure from scratch.

Yes. Chat with your data is built in: each organization gets a private knowledge base where you upload PDF, DOCX, Markdown, or TXT files (up to 10 MB each) and the assistant answers from them through a searchKnowledgeBase tool that cites its sources in chat. Retrieval runs on pgvector with embeddings scoped per organization, so one tenant's documents never appear in another's answers.