Concepts

The AI package is organized around four domains:

• agents: model, instructions, tools, and delegation
• tools: canonical tool definitions and transport-specific resolvers
• chat: server streaming and client UI primitives
• mcp: content helpers for model context protocol tools

Why this split exists

AI applications quickly become hard to maintain when model prompts, database mutations, UI rendering, and external surfaces all share one ad hoc abstraction.

Tulip AI keeps the boundaries explicit:

• define tool behavior once
• resolve it differently for agents, MCP, and UI metadata
• keep server code server-only
• keep client tool renderers type-safe without bundling server code
• let agents delegate to specialist agents instead of growing one giant prompt

Recommended architecture

Use one orchestrator agent for the workspace and specialist agents for bounded domains:

• Sales: CRM, prospects, duplicate checks, qualification
• Delivery: projects, initiatives, tasks, planning
• Operations: time analytics, initiative health, effort-vs-scope
• Knowledge: Drive, folders, documentation lookup

This keeps tool choice and instructions easier to reason about.