Model Context Protocol · stdio

Hand your agent the whole library

Connect Frontary once and your assistant can search it, compare moods, and pull the real prompt for anything you are entitled to — without you opening a browser and copying anything by hand.

See the plans Read the docs

Six tools

What your agent can ask

search_assets

Find by mood, tag, family or plain words. Returns the same cards the library shows, with a page URL for each.

list_families

Every family and how many published assets it holds, so a query is never guessed at.

get_asset

One asset in full: description, the controls it shipped with, its stack, licence, delivery options and which plans cover it.

compare_assets

Two to four assets side by side on the fields that decide a purchase.

get_prompt

The build prompt, delivered by the store. If you are not entitled, the store's own refusal comes back in its own words.

list_plans_and_bundles

What can actually be bought, priced for your region.

Connect it

Two lines in your MCP config

The server is a single file with no dependencies. Point your client at it and it reads the same store this site does — 23 published assets across 9 families.

It is read-only. Nothing here writes to the store; that is Admin's job and it stays behind a session.
It does not decide access. Every entitlement answer comes from the Worker, so a plan change in Admin shows up on the next call.
Pass FRONTARY_TOKEN with a signed-in buyer's token and get_prompt will deliver; leave it out and it will tell you, in the store's words, what is missing.

{
  "mcpServers": {
    "frontary": {
      "command": "node",
      "args": ["path/to/frontary/mcp/server.mjs"]
    }
  }
}

Questions

Before you wire it in

Does the agent get to see assets I cannot open?

It can see that they exist, because the library page shows them too. It cannot take a prompt or a source for one your session is not entitled to — the store refuses and the refusal is passed through unchanged.

Why stdio rather than a hosted URL?

Because a hosted endpoint is a second front door to the store and would need its own access policy before it could exist. A local process uses the same API the browser does and inherits whatever the buyer's own credentials allow.

Will it invent a control or a price if my agent asks badly?

No. The tools return the fields the store publishes and nothing else, and the server's instructions tell the client not to make up a licence term, a price or a control these tools did not return.

What if the store is unreachable?

The tool says so, in the result. A silently empty list would read to an agent as “Frontary has nothing about that”, which is a worse failure than an honest one.