Curacity Brain

Your company, all in one place

Ask me anything about Curacity — products, processes, pricing, clients, team, or policies. I pull from all connected sources in real time.

Coda FAQs & Docs

Google Drive

HubSpot CRM

Slack

Notion

Asana

Gmail

Google Calendar

Live sources

Watch which systems are queried for each answer

Ask a question to see which sources are queried in real time

How Curacity Brain works

One Claude API call, seven connected data sources, and a dynamically built system prompt. Here's the full picture.

👤

Employee

asks a question

→

🔄

MCP servers

live context pull

→

📝

System prompt

assembled dynamically

→

🧠

Claude

claude-sonnet-4

MCP servers fetch live context before every response

Instead of a static system prompt, this bot queries your actual live tools — Coda, Google Drive, HubSpot, Notion, Asana — at request time via MCP (Model Context Protocol) servers. Claude reads the fresh data and synthesizes an answer grounded in what's actually true today.

// MCP servers passed to each API call
mcp_servers: [
  { type: "url", url: "https://coda.io/apis/mcp", name: "coda" },
  { type: "url", url: "https://mcp.notion.com/mcp", name: "notion" },
  { type: "url", url: "https://mcp.hubspot.com/anthropic", name: "hubspot" },
  { type: "url", url: "https://mcp.asana.com/v2/mcp", name: "asana" },
  { type: "url", url: "https://gcal.mcp.claude.com/mcp", name: "gcal" }
]

The system prompt is your company's identity, hard-coded once

We pre-seeded the system prompt with everything we learned about Curacity from scanning your sources: what VISTA Core and VISTA AI are, how attribution works, your pricing model, your policies. This is the "always true" layer. The MCP tools add the "latest" layer on top of it at runtime.

Claude decides which tool to call based on the question

When someone asks about hotel clients, Claude searches HubSpot. FAQs → Coda. Policies → Google Drive. Active sprint → Coda Product hub. Projects → Asana. Claude reasons about the question and picks the right source. It can call multiple sources for complex questions and synthesize them together.

// The API call structure
const res = await fetch("https://api.anthropic.com/v1/messages", {
  method: "POST",
  body: JSON.stringify({
    model: "claude-sonnet-4-20250514",
    max_tokens: 2000,
    system: CURACITY_SYSTEM_PROMPT,
    messages: conversationHistory,
    mcp_servers: MCP_SERVERS    // ← this is the magic
  })
});

Conversation history = memory across turns

Claude has no persistent memory by default. We maintain a history array in the browser and send the full thread with each call, so follow-up questions work naturally: "tell me more about that" or "what's our policy on the second point you mentioned" all resolve correctly.

For production: proxy the API call through your backend

This demo calls the Anthropic API directly from the browser. In production, route the call through a backend endpoint (Node.js, Python, etc.) so your API key stays server-side and you can add auth, rate limits, and logging.