Key takeaways
- MCP, the Model Context Protocol, is an open protocol that lets an AI assistant such as Claude call tools and read data in another system through one standard connection.
- A connected assistant can only do what the connection’s key scopes allow; the assistant gets no separate, broader access to the store.
- On Torva, Growth connects Claude for reading, and Command adds storefront and email drafts that nothing publishes or sends until you do.
- Use one key per client, grant the fewest scopes that do the job, and revoke the key to disconnect.
- AI-written storefront or email copy still has to avoid health, dosing and drug-brand claims, so every draft goes through the same review as copy you write yourself.
What MCP is, in one paragraph
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant connect to tools and data in another system. The system publishes an MCP server listing named tools, the assistant connects with credentials you provide, and when you ask a question the assistant decides which tools to call.
The important part for a store owner is the boundary. The assistant does not get a login to your admin. It gets exactly the tools the server offers and exactly the permissions of the key behind the connection. On Torva, a key that cannot read orders through the API cannot read them through Claude either, because it is the same check.
What Claude can do with your store
Once connected, Claude can answer questions about your store from live data and prepare drafts for you to review. It is most useful for questions that would otherwise mean exporting a spreadsheet.
- “Which ten products had the most orders last month, and how many units of each are available now?”
- “List variants with fewer than 20 available units, sorted by how fast they sold in the last 30 days.”
- “How many orders are unpaid or unfulfilled, and which are older than three days?”
- “Read my existing emails and draft a restock notice in the same voice for the 10mg kit.” (Command)
- “Rewrite the About page draft to be shorter and plainer, keep the research-use statement word for word.” (Command)
- “Show me any content review findings on the storefront and suggest replacement wording.”
Permissions: who can do what
Access is set by plan and by key scopes. Growth connects Claude for reading: catalog, orders, inventory, store settings and emails. Command adds writing storefront and email drafts. Nothing is published or sent without you.
Writes need a live key with the matching scope and full API access for the store; a test key cannot write at all. Revoking the key disconnects every client using it.
| Tool | What it does | Read or write | Plan |
|---|---|---|---|
| get_store | Store id, customer-facing name, currency, open/draft/closed | Read | Growth |
| list_products | Products with id, title, slug, status, last change | Read | Growth |
| list_orders | Orders with totals, payment and fulfilment status | Read | Growth |
| get_inventory | On hand, reserved and available per variant | Read | Growth |
| list_emails | Automations and templates with full content | Read | Growth |
| list_content_reviews | Findings from Torva’s review of your store content | Read | Growth |
| get_storefront | The saved storefront draft and its version | Read | Growth |
| write_storefront | Saves a storefront draft for review; never publishes | Write (draft) | Command |
| write_email | Creates or revises an email draft; never sends | Write (draft) | Command |
| rename_store | Changes the business name customers see | Write (direct) | Command, settings.write scope |
What the tools cannot do
Product titles, prices, stock and descriptions are catalog records and cannot be changed through write_storefront. Email campaigns cannot be written through write_email. Reads return 25 records at a time, so Claude pages through larger lists with an offset.
Connect Claude step by step
You need a Growth or Command plan and your store’s connection URL, which the Desk shows under Settings → Developers → AI connections.
- In Claude, open Settings → Connectors.
- Choose Add custom connector.
- Paste https://torvaos.com/api/mcp/<your-store-id>, using the URL from the Desk.
- Choose Connect.
- Sign in to Torva when prompted.
- Review the access requested and choose Allow.
- Start a new chat and ask a simple read question, such as “What is my store’s name and currency?”, to confirm it works.
Claude Code or Codex
Coding agents connect as an HTTP MCP server using a Torva API key as a Bearer token. Create a key in the Desk under Settings → Developers with only the scopes the agent needs.
In Claude Code, keep the key in an environment variable and add the server with: claude mcp add --transport http torva https://torvaos.com/api/mcp/<your-store-id> --header "Authorization: Bearer $TORVA_API_KEY". In Codex, add an HTTP MCP server with the same URL in its configuration and supply the key as a Bearer token from an environment variable; check Codex’s current MCP documentation for the exact field names.
Keep the secret in the client’s credential settings or environment, never in a project file that is committed, shared or pasted into a chat.
Safety practices
Treat a connected assistant like a new colleague with a key to part of the building: useful, fast and checked. These habits keep the risk small.
- Least privilege: start with read scopes; add draft writing only when you want drafts.
- Separate keys per client: one for Claude, one for each coding agent, so you can revoke one without breaking the others.
- Leave settings.write off unless you intend Claude to rename the business.
- Review every draft in the Desk before publishing or sending.
- Protect customer data: do not paste customer names, addresses or order details into unrelated chats or tools.
- Audit: review active keys monthly and revoke any you no longer use.
Troubleshooting and compliance
Most connection problems have one of three causes, and all three show up as clear errors.
- 401 Unauthorized: the connection was added but not completed. Choose Connect, sign in to Torva and allow access.
- A tool call is refused: the key lacks that scope, the plan does not include it, or it is a test key trying to write.
- Answers seem incomplete: reads return 25 records per call; ask Claude to page through all results before summarizing.
- stale_version on a storefront draft: someone saved a newer draft; Claude should read it again and reapply only the requested change.
AI copy still has to pass review
Copy Claude writes carries the same rules as copy you write: no health or efficacy claims, no dosing or human-use language and no prescription drug brand names. That is why Torva’s write tools produce drafts, not published pages. Ask Claude to call list_content_reviews after editing, and read our guide to product descriptions without claims.
How Torva handles this
Torva’s MCP server exposes the same actions as its developer API, with the same authentication and scope checks, so the assistant never has a second route into your store. Growth includes the supported read connection; Command adds storefront and email draft editing, full API access and webhooks. See developers for keys and scopes.
If you would rather not connect an outside assistant at all, Ask Torva is the assistant built into the Desk. Plans on pricing apply when public activation opens; you can preview a store free now, no card needed.
Frequently asked questions
What is MCP?
MCP, the Model Context Protocol, is an open protocol for connecting AI assistants to external tools and data. A service publishes an MCP server describing what it can do, such as listing orders or reading inventory, and an assistant like Claude connects to it and calls those tools when your question needs them. Access is set by the credentials you give the connection, not by the assistant.
Can Claude change my store or send emails?
On Torva, only within the scopes you grant and never all the way to customers. With Growth, Claude can read the catalog, orders, inventory, store settings and emails. Command adds saving storefront drafts and email drafts. A draft is never published or sent by Claude; a person reviews it in the Desk. The one direct write is renaming the business, which needs its own scope you can leave off.
Why does Claude say it is not authorized when I use the store connector?
A 401 response usually means the connection has not been completed: open the connector in Claude, choose Connect, sign in to Torva and allow access. If Claude connects but a specific action is refused, the key behind the connection lacks that scope, or it is a test key, which cannot write. Adjust scopes in the Desk under Settings, Developers, then reconnect.
Is it safe to connect an AI assistant to my store?
It can be, with least privilege. Give each client its own key, grant only the scopes it needs, keep the key in the client’s credential settings rather than in a chat, and revoke it when you stop using it. Review drafts before publishing, and avoid pasting customer personal data into unrelated conversations. Treat the assistant as a fast colleague whose work you check.
Preview your store on Torva OS — no card needed — or explore the Desk demo, the eight themes and the plans. Questions: hello@torvaos.com.
General information for merchants, not legal, tax or financial advice. Check the rules in your jurisdiction and with your providers.