> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryhoard.com/llms.txt
> Use this file to discover all available pages before exploring further.

# What your AI assistant can do

> An honest map of what a connected AI assistant (Claude, ChatGPT, Codex) can read and change in your Hoard store today, and what's still dashboard-only.

Once you've [connected Claude](/connect-to-claude), [ChatGPT](/connect-to-chatgpt), or [Codex](/connect-to-codex), the assistant gets a scoped read of your Hoard store and can run carefully-gated changes. This page is the honest map: what's exposed today, what's not, and how to set expectations when you ask a question.

## What an assistant can read

These are safe to ask any time. The assistant pulls live data from your account; nothing changes.

**Your inventory at a glance**

* "What's my total listed inventory value?"
* "Break down my listed value by game."
* "How many cards do I have, and how many are flagged as underpriced?"
* "Walk me through my top 10 most valuable cards."
* "Show me cards with the biggest gap between my listed price and market."

**Pricing rules**

* "What pricing rules do I have active?"
* "Which rule covers the most cards?"
* "Preview what would happen if I bumped my Mythic rule from 1.20x to 1.25x."

**Movers and price changes**

* "Show me my biggest gainers in the last 7 days."
* "Any cards in my store flagged with 'spike' volatility?"
* "What were my biggest price losers this week?"

**Sales and revenue**

* "What was my revenue last month?"
* "How are fees trending? Has TCGplayer's rate moved?"
* "Show me my last 3 months of orders + net-after-fees."

**Customers**

* "Who are my repeat customers?"
* "What percentage of my revenue comes from my top 10 buyers?"
* "Has anyone bought from me more than twice?"

**Sync status**

* "When did Hoard last sync?"
* "Is a sync in progress right now?"

**Smart Actions**

* "What suggestions should I review today?"
* "Show me stale-stock sale candidates."
* "Which product lines have pricing-rule scope drift?"

## What an assistant can change (with permission)

Hoard treats every write action through a **preview → confirm → commit** flow with server-side safety checks. The assistant never directly mutates your store on a single call — it always shows you what would happen first.

**Today:**

* **Edit a single pricing rule.** Change the multiplier, toggle never-go-down, switch the price source. The assistant calls preview first, you see the projected impact + risk classification, and (depending on your permission settings) either auto-commits low-risk edits or asks you to click a confirmation link.
* **Accept or dismiss Smart Actions.** The assistant can list recommendations, accept one into its workflow, or mark it not useful. Price-related recommendations still queue a price-update preview; accepting one does not silently push prices to TCGplayer.
* **Trigger a sync.** "Trigger a sync" is a write action, not a read — it tells Hoard to go pull fresh data and recompute. Asking when Hoard last synced, or whether one is running, is read-only; actually kicking one off is on the write side.

**Coming soon (filed in the roadmap, not shipped yet):**

* Mass reprice through the same plan-token flow
* Per-card price overrides
* Running a discount sale
* Toggling pricing rules on or off

See [Agent permissions and safety](/agent-permissions) for how the gating works.

## What an assistant can't do today

The honest list. These are dashboard-only as of this writing (last reviewed June 2026):

* **Adding new inventory.** Quick Add, voice match, paste-list signup flow — all dashboard-only.
* **Changing any account settings.** Password, timezone, notification preferences, pricing template, product line selection, shipping adder — dashboard-only.
* **Managing pull sessions.** The assistant can start a session, but viewing it, picking through it, deferring, or completing — all dashboard-only.
* **Manual per-card price overrides.** Lock a price, clear a lock — dashboard-only.
* **Consignment.** Creating consignors, assigning cards, payouts, share links — dashboard-only.
* **Triggering or downloading rollbacks.** Dashboard-only.
* **CSV imports or exports.** Dashboard-only.

This isn't because we couldn't expose them — it's that we want the AI surface to start with the read-mostly + heavily-gated write path before opening up everything. Each new capability gets the same preview→confirm→commit + risk-gating + audit log treatment.

## How an assistant figures out what to do

Under the hood, Hoard exposes a small set of tools to your assistant. There are four fast lanes for common reads (`hoard_list_inventory`, `hoard_movers`, `hoard_market_signals`, `hoard_sync_health`) that take arguments and return JSON directly. There's `hoard_search` for discovering what else is available. And there's a read/write split for everything else: `hoard_read` runs a small JavaScript program against your store with write actions blocked, while `hoard_write` runs the same kind of program with changes allowed (and still gated behind the preview→confirm→commit flow above). An older `hoard_execute` alias still forwards to `hoard_write` for legacy setups.

The read/write split is the important part: anything that only looks at your data runs through `hoard_read`, so a read-only question can never accidentally change something. This "code mode" lets the assistant compose multiple data fetches in a single round trip instead of taking one tool call per question.

You don't need to know any of that to use it. You ask in plain English; the assistant figures out the JS. But it explains why the same question sometimes feels instant (one call) and sometimes feels chunkier (multiple calls when the question needs cross-referencing data).

## Setting expectations: "I sell Lorcana / One Piece / Riftbound"

Claude and other large models have deep knowledge about Magic: The Gathering and decent knowledge about Pokemon. Knowledge of newer games (Lorcana, One Piece, Riftbound, Flesh & Blood, Star Wars Unlimited, Digimon, Pokemon Japan) is thinner — most of these games didn't exist or weren't well-covered when the models were trained.

What this means in practice: your assistant can reliably reason about your **actual data** for any game (your prices, your movers, your sales, your rules), but it should ask before making strong claims about market norms for games where its training is thin. If you're a Lorcana-focused seller and your assistant says "Lorcana commons typically run \$0.10–0.50," treat that as a guess unless it's backed by your own data.

A pattern that works well: **ground your assistant in your own inventory** by starting a chat with something like "I sell mostly Lorcana, here's a snapshot of my store" and let it run the read-side queries first. Reasoning improves a lot once it's calibrated to your data instead of generic priors.

## Revoking access

Open [Settings → Assistants](https://www.tryhoard.com/settings/connected-apps). Click **Revoke** next to whichever assistant you want to disconnect. The assistant immediately loses access — it can't refresh its way back in. Any in-flight preview that hadn't been committed is also dead.

You can re-connect any time by following the [connect guide](/connect-to-claude) again.

## Tutorials

Walk-throughs for the flows sellers actually run into:

* [Your first agent-driven price change](/tutorials/first-agent-price-change)
* [Picking the right permission mode](/tutorials/picking-a-permission-mode)
* [Reading your agent activity log](/tutorials/reading-the-audit-log)
* [When the agent says 'I can't do that'](/tutorials/when-the-agent-says-no)
* [Rotating your agent bearer key](/tutorials/rotating-your-bearer-key)
* [Going from cautious to confident](/tutorials/trust-progression)

## Related reading

* [Connect Hoard to Claude](/connect-to-claude)
* [Connect Hoard to ChatGPT](/connect-to-chatgpt)
* [Connect Hoard to Codex](/connect-to-codex)
* [Agent permissions and safety](/agent-permissions) — how Hoard keeps an AI assistant from doing something dangerous
* [Security](/security) — the broader trust model