The infrastructure for AI-powered shopping is maturing faster than most people realize. Skills, agents, and the Model Context Protocol (MCP) are creating a new layer between "I want to find a deal" and "I found the best deal available right now." Here's how the stack works — and where Closetta fits into it.
The Three Layers of AI Shopping
Layer 1: The Data Source
An AI agent is only as good as the data it can access. For fashion deal-hunting in Canada, that means a source that:
- Updates daily (not cached from training)
- Covers a wide enough set of brands to be useful (67+ in Closetta's case)
- Includes historical context, not just current status
- Is machine-readable — structured JSON, not HTML that needs to be scraped ad hoc
Closetta operates at this layer. The public API and MCP server expose sale data in formats designed for agents to consume.
Layer 2: The Reasoning Engine
The language model (Claude, GPT-4, Gemini) sits here. It interprets user intent, decides which tools to call, synthesizes the results, and generates a response that's actually useful.
A good reasoning layer does things like:
- Understand "is this a good deal?" requires both current price AND historical context
- Know when to call
search_sales_by_brandvs.get_active_salesvs.get_trends - Apply judgment: "40% off Lululemon is significant because they almost never discount; 40% off RW&CO is below their average"
The model doesn't store deal data — it calls the tool that has it.
Layer 3: The Skill or Agent Wrapper
This is the glue. An AI "skill" is a pre-configured agent with specific tools, instructions, and a defined purpose. Instead of configuring each session from scratch, a user invokes a skill and gets an assistant that already knows its job.
A Closetta shopping skill would have:
- Access to the MCP tools (
get_active_sales,get_trends,search_sales_by_brand) - Instructions for interpreting discount data in context
- Persona: a knowledgeable friend who obsessively tracks Canadian fashion sales
- Output format: direct, data-grounded answers — no marketing fluff
When someone asks "should I buy from Rudsak today?", the skill calls search_sales_by_brand("Rudsak") and get_trends, compares the current discount against the historical peak, and gives a specific recommendation: "Rudsak peaked at 50% in February 2026. They're currently at 30%. Not the best window unless you need it now."
What MCP Changes About Agent Design
Before MCP, connecting an AI agent to external data required custom code: an API wrapper, prompt injection, response parsing. Every tool integration was one-off work.
MCP standardizes this. A data source publishes a schema that describes its tools — inputs, outputs, descriptions. An agent connects to the endpoint. The model can read the tool descriptions and know how to use them without any additional code.
For Closetta:
{
"mcpServers": {
"closetta": {
"url": "https://closetta.app/api/mcp"
}
}
}
That single config entry gives any MCP-compatible agent access to 67 brands' worth of live sale data, plus 5+ months of historical trends. The agent can answer questions that no amount of training data could answer — because training data is static, and sale windows last days.
Agentic Workflows for Fashion Shopping
Here's what a fully agentic shopping workflow looks like in 2026:
Trigger: User says to their AI assistant: "I need new running shoes before my half-marathon in July. What's the best time to buy and what brands should I look at?"
Agent step 1: Identify relevant brands → call get_brands, filter by category "Sportswear" and "Footwear"
Agent step 2: Check current sales → call get_active_sales, filter for sportswear/footwear brands
Agent step 3: Find historical peaks → call get_trends, look at Nike, Adidas, New Balance, Asics, Skechers across months
Agent step 4: Synthesize and respond:
"For running shoes in Canada, here's what the data shows: Nike and Adidas currently have no active sales — their 2026 events have been restrained at 20–30%. New Balance appeared at 25% in May (their first tracked event this year). Asics and Skechers have been more aggressive, hitting 40–50% during clearance windows. If you have until July, set an alert for Asics and Skechers — their next clearance window could hit 40%+. If you need shoes now, check Skechers at [current discount]."
That answer required four tool calls, cross-referencing live data against historical trends, and category-aware reasoning. A human could do this by spending 20 minutes checking each brand's site. An agent does it in seconds.
Skills vs. General-Purpose Agents
A general-purpose AI assistant (Claude without specific configuration) can answer shopping questions if it has access to Closetta via MCP. But it doesn't know it should check historical context before giving a recommendation. It might answer with training knowledge rather than live data.
A shopping skill (Claude + Closetta MCP + specific instructions) is pre-configured to always:
- Query live data first
- Apply historical context before making a recommendation
- Quantify rather than qualify ("60% off" beats "major discount")
- Flag when a current sale is above or below the brand's typical depth
The skill is the difference between a good answer and a reliably good answer.
What AI Agents Can't Do Yet
For completeness: here's where the current stack falls short.
Item-level data. Closetta tracks brand-level discounts, not specific products. An agent can tell you The North Face is at 40% off; it can't tell you whether the specific jacket you want is included.
Inventory. No real-time stock data. The agent can identify that a sale is running but can't tell you whether XS is still available.
Cross-retailer comparison. If the same jacket is sold by The North Face directly and also by another retailer, Closetta only sees the brand's own site.
Purchase execution. Today, the agent surfaces the deal. You still click the link and buy. Agentic purchasing (with explicit user authorization) is the next evolution.
These gaps are real but don't undermine the core value: knowing which brands are on sale, how deep the discount is, and whether that's historically unusual — which is the information that most shoppers lack when making a buying decision.
For Developers: Adding Closetta to Your Agent
MCP endpoint: https://closetta.app/api/mcp
REST endpoints:
GET /api/v1/sales— live active and recently ended salesGET /api/v1/brands— all 67 tracked brandsGET /api/v1/trends— monthly discount historyGET /api/v1/brand-history— history for a specific brand
No authentication. CORS enabled. Cached at CDN, refreshed hourly.
If you're building a shopping assistant and want real-time Canadian fashion data, Closetta is designed to slot into that stack directly. The MCP server handles tool discovery automatically — the model reads the tool descriptions and knows how to use them.
The Broader Shift
The pattern here — data source + reasoning layer + skill wrapper — applies beyond fashion. Price tracking for electronics, inventory monitoring for secondhand goods, event calendar aggregation for travel. Every domain where real-time, structured data exists and users benefit from AI reasoning on top of it.
Closetta was built for Canadian fashion specifically because the domain has clear pain points: too many brands, inconsistent sale calendars, and marketing language that obscures whether a deal is actually good. The data infrastructure (daily scraping, historical tracking, machine-readable APIs) is what makes AI-grounded answers possible.
Shoppers can use the price tracker directly. Developers and agents can use the MCP server or the REST API. Both paths lead to the same underlying data — 67 brands, updated daily, with enough history to know whether today's sale is worth acting on.
Discount data sourced from Closetta's daily AI monitoring across 67 brands. Historical patterns reflect observed trends and are not guarantees of future sale events.