Skelet 0.3.1

AI agents can read the store over MCP. /api/admin/mcp serves the dashboard chat’s tools to any MCP client that can send an integration token in X-Api-Token, Claude Code included. It’s read-only, reads one channel per connection, and offers only the tools the token’s permissions cover - sales, catalog, orders, stock, promotions, reviews and store setup. Each token can run 4 calls at once and, with RATE_LIMIT_ENABLED on, MCP_RATE_LIMIT_PER_MIN a minute. MCP_ENABLED=false turns it off. Setup is in docs/mcp.md.

The storefront packages publish to npm. @skelet/storefront-core, -next, -astro and -react go out with every release at its version, compiled to JavaScript with their types and licensed MIT, so a shop in a repo of its own can install them. A Skelet checkout keeps reading the source. The release workflow publishes through npm’s trusted publishing, with no token stored - docs/releasing.md has the one-time setup.

The AI chat’s product search reaches the whole catalog. It used to scan only the first thousand products, so on a bigger store it answered “no products match” for things you sell. It now searches in the database, by name, SKU or URL, like the Products list does, newest first.

The AI chat’s tools answer for the chat’s channel. A product’s review counts and stock by location covered every channel, and review counts with the dashboard set to “All channels” did too.

A failed read in the AI chat says so. “Could not load …” came back to the model as an ordinary answer, and the cause was never logged. It’s flagged as an error now, and logged.

The AI chat’s overview tools say when a read failed. Sales, web traffic, customer behaviour and experiments answered “no data” when the database or the Signals service couldn’t be read, and the chat’s opening snapshot left the section out the same way. They flag the failure now, and the snapshot says the section couldn’t be read.

Money in the AI chat uses each currency’s own decimals. Every amount was printed with two, so yen came out a hundred times too small and dinars ten times too large. Promotion totals now name their currency too.

The AI chat’s low-stock list covers the channel’s own warehouses. It filtered a store-wide page of 500 items down to the channel’s, so a busy store could show “no low-stock items” for a channel that had some, and in flat inventory mode it never showed any. The insights card’s count is the channel’s now too.

The AI chat’s window tops out at a year. A larger window reached the sales queries unclamped.

The AI chat can only run the tools it was offered. With customer data turned off in the AI settings, the chat hid the customer-record tools, but the model could still run one by asking for it by name. Now any tool it wasn’t offered is refused.

Turning off customer data in the AI settings hides emails on orders too. The order tools kept showing each customer’s email after the switch was off. They don’t now.

The AI chat quotes what shoppers and visitors wrote. Customer names, review summaries, site searches, page addresses, referrers, rage-clicked element text, JS errors and custom event names reach the model in quotes, with line breaks escaped, so a crafted name or search can’t pass itself off as part of the tool’s answer.

The AI chat no longer confirms an id belongs to another channel. Looking up an order or customer from another channel answered differently from one that doesn’t exist; both read the same now.