Personalised Recommendations Agent

Expose user insights and catalogue as MCP tools

Create MCP tools that give the agent access to user behaviour signals and your product/content catalogue.

Tools to expose via MCP:
- get_user_profile(user_id) → preferences, segments, history
- get_recent_activity(user_id, days?) → recent views, purchases, clicks
- search_catalogue(query, filters?) → products/content with metadata
- get_item_details(item_id) → full spec, pricing, related items

Define what 'recommendation' means in your context

The MCP server knows your domain. You define what data to expose — product SKUs, content IDs, course modules, subscription tiers. The AI reasons over whatever you provide.

Register and connect in SentientOne

Add the MCP server in the SentientOne dashboard. The agent will call your tools dynamically as needed to build a full picture before generating recommendations.

Create the Recommendations Agent

Define an agent with a system prompt that establishes recommendation principles for your domain.

System Prompt:
You are a personalised recommendation engine for [Platform Name].
Given a user's profile and recent activity, recommend [products/content/courses].
Rules:
- Never recommend items the user already purchased
- Prioritise items that match their explicitly stated preferences
- If a user is mid-journey (e.g. completed beginner content), recommend the logical next step
- Return exactly 3–5 recommendations, ranked by relevance
- For each recommendation, include: item ID, name, one-sentence reason why

Attach the insights MCP tools

Connect your MCP server. The agent will automatically use get_user_profile and get_recent_activity to gather context before searching the catalogue for matches.

Configure response format

Set up the agent to return structured JSON responses so your application can directly render the recommendations without additional parsing.

Call the agent with user context

Pass the user ID and a trigger event to the agent. Let it fetch the full profile via MCP:

const response = await fetch("https://app.sentientone.ai/api/chat", {
  method: "POST",
  headers: {
    "X-Agent-Id": "YOUR_RECOMMENDATIONS_AGENT_ID",
    "X-App-Key": "YOUR_APP_KEY"
  },
  body: JSON.stringify({
    message: `User ${userId} just completed action: ${trigger}. 
               Generate personalised recommendations.`,
    sessionId: userId
  })
});
const { reply } = await response.json();
// reply contains structured recommendations

Trigger from key moments

Call the agent on purchase completion, course completion, session start, or weekly email send. Each trigger produces fresh, context-aware recommendations based on the latest user data.

Render in your UI or email template

Parse the structured JSON response and render the recommended items in your product detail page, home feed, or transactional email. No AI code in your rendering layer — just data.