I gave an AI agent the keys to a store

Most software in the AI-commerce space is a dashboard. You log in, you click through tabs, you read a score, you press some buttons. I built Pollen differently. Pollen is a thing an AI agent can operate. You connect it to an AI client like Claude, and you say “audit my store’s AI visibility and fix what you can,” and it does the work. This is a note on how that is built, what an agent can and cannot do through it, and why I made that bet.

The bet

The people who worry about whether ChatGPT or Gemini recommends their products are, increasingly, the same people who already spend their day inside an AI client. If that is where your buyer works, that is where your product should be operable. A dashboard asks them to leave that context and learn your tabs. A well-built server lets them stay where they are and ask for the outcome.

So Pollen ships as a Model Context Protocol (MCP) server. MCP is the open standard, originally from Anthropic and now consumed by OpenAI, Google, and Microsoft too, that lets an AI client call tools on a server. It is the same idea as an API, with one difference that matters here: the caller is a language model reasoning on the user’s behalf, not a script a developer wrote in advance.

What an agent can actually do

Through Pollen’s management server, an agent working for a merchant can run the whole loop. Concretely, the tools cover:

• See where you stand. Pull the account summary and the agent-readiness score, list products, and read exactly which gaps cost points.
• Sync and enrich. Sync the catalog from Shopify or WooCommerce, then run Pollen’s multimodal enrichment, which reads the product text and the product photos together and returns structured facts, an agent-ready description, and search terms.
• Fix, carefully. Apply the automated fixes: fill empty SEO titles and meta descriptions, add tags, write image alt text, set a missing brand. These are fill-only-empty against the live store, so a value you set yourself is never overwritten.
• Supply what only you know. Tell it “the tote bag weighs 200 grams” and it writes that fact to your store through Pollen.
• Publish. Push the enriched catalog to Google Merchant Center, Meta, and the store’s own feeds, and check pipeline status.

None of this is a demo mode. It is the same code paths the dashboard uses, exposed as tools with titles and behavior annotations so the client knows which ones only read and which ones write.

Three surfaces, three audiences

“An MCP server” is actually three servers in Pollen, because three different kinds of agent need three different things.

The management server is per-merchant and authenticated: the agent acts as you, so it uses OAuth or an API key. The storefront catalog server is public and read-only, so a shopper’s agent can search a store’s products without any account. The network server is the early, cross-store surface, and it is consent-gated by design.

The honest limits

Agent-operable does not mean unattended. A few lines I drew on purpose:

• The OAuth click stays human. Connecting a store, or granting Google and Meta access, opens a consent window the person approves. An agent can walk you to the door; it does not sign for you.
• Writes are annotated and conservative. Every tool declares whether it only reads or can change the store, and the fixes never overwrite something you wrote. Replacing a product description is never a silent batch operation; it is per product and merchant-approved.
• Agents still misread artifacts. Early on, a client quoted Pollen’s internal enrichment draft as if it were the text that would be published. Pollen now returns a clearly labeled, sanitized “this is what would be written” candidate next to the raw draft, because an agent will confidently narrate whatever you hand it.

Why I think it is worth it

The reason Pollen exists at all is a gap I can measure. I asked ChatGPT, Gemini, and Perplexity fifteen ordinary shopper questions about products a real connected store sells. Across three assistants, zero of forty-five answers named the store. Competitors were named instead. Being online is not the same as being recommended, and closing that gap is tedious, per-product data work.

Tedious, per-product data work is exactly what an agent is good at, if you give it the right tools and honest guardrails. That is the whole idea: put the fix where the buyer already is, let the agent do the grind, and keep the human on the decisions that need judgment.

Pollen is live on WooCommerce today, with a Shopify listing in review. The servers are listed on the official MCP Registry (in.elytron/pollen and in.elytron/pollen-store), and the docs are at github.com/elytronlabs/pollen-mcp. If you run a store and want an agent to keep it AI-ready, start at pollen.elytron.in.