An embedded Salesive App can give Ola tools she is allowed to call. The app registers each tool with modelContext.registerTool, the dashboard passes the descriptor to Ola, and when she calls one the merchant sees an approval card with Approve, Reject and Always allow. Nothing runs until the merchant chooses.
What is WebMCP in this context?
WebMCP is a way for an embedded page to describe the actions it can perform so an AI assistant can use them. In Salesive, the page is the app running inside the dashboard. The assistant is Ola. The tool descriptor is how the page says, in effect, 'this is what I can do and this is what it is called'.
The important part is the direction of control. The app does not reach into Ola, and Ola does not get a new permission. The app publishes a menu of actions it already supports, and the merchant decides whether Ola may use one. It is closer to giving an assistant a button than to giving it a key.
How does an app register a tool?
The embedded app calls modelContext.registerTool with a descriptor for each action it wants to expose. The dashboard forwards those descriptors to Ola, so she knows the tool exists and what it is for. Registration is a description step, not a grant: it does not give the app any new access.
Because the app is embedded, it is already running in the merchant's dashboard session. That is why no OAuth scopes are involved and no server access is needed. The tool runs in the app's iframe, in the same page context the merchant can already see. If the app is not open, there is nothing to call.
What does the merchant see?
When Ola calls a registered tool, the merchant sees an approval card. It offers three choices: Approve, Reject and Always allow. Approve runs that one call. Reject stops it. Always allow lets that tool run without asking again.
That card is the whole safety model, and it is a good one. The merchant is not asked to trust an invisible agent. They see the action at the moment it matters, with the option to stop it. Always allow is there for tools the merchant trusts and uses often, so the approval does not become a reflex that everyone clicks without reading.
What can a tool actually do?
A tool can only do what the page can already do. It gets no extra privileges, no server access and no scope that the app did not already have. If the app cannot delete an order from its own interface, a tool cannot delete an order either. The tool is a shortcut to an existing capability, not a new one.
That constraint is what makes WebMCP safe to offer. The worst a malicious tool can do is the worst the app could already do in its own iframe, in front of the merchant. The approval card adds a checkpoint, and the iframe keeps the action inside the page the merchant installed.
Handle failure visibly. If a tool cannot complete, return a clear error to the page and show it in the app's interface, so the merchant sees what happened rather than a spinner that never stops. Ola can report that the call failed, but the app still owns the explanation.
How should you design a tool?
Name the tool for the job, not for the implementation. 'Create invoice' is easier for Ola to choose correctly than 'runInvoiceJob'. Describe what it does and what it changes in plain language, because that description is what Ola uses to decide when to call it. If the tool writes data, say so in the description.
Keep tools narrow. One tool that creates an invoice is easier to approve than one tool that 'manages billing' and does five different things depending on the input. Narrow tools also produce better approval cards, because the merchant can see exactly what they are agreeing to. Test the tool through Ola, not only through your own interface, since the descriptor is what she sees.
When should you use WebMCP instead of the API?
Use WebMCP for actions that belong to the app's interface and need the merchant in the loop: creating an invoice, applying a setting, running a report the merchant is watching. Use the API for background work that happens without a page open, such as syncing orders or reacting to a webhook.
The two can work together. A tool can be the trigger that starts an action in the app, and the app can then use its scoped token to call the API for the parts that need server access. That keeps the merchant's approval at the front and the heavy work where it belongs. If you have not built the app yet, start with what Salesive apps are.
A tool should be a button the merchant can see, not a permission the merchant has to imagine.
Frequently asked questions
Does WebMCP need OAuth scopes?
No. A registered tool runs in the app's iframe and can only do what the page can already do. There are no extra scopes and no server access involved in the tool call itself.
Can Ola call a tool without approval?
Only if the merchant chooses Always allow for that tool. Otherwise each call shows an approval card with Approve, Reject and Always allow, and nothing runs until the merchant picks one.
Does the tool run on Salesive servers?
No. The call runs in the app's iframe, in the dashboard the merchant already has open. If the app is not open, there is nothing for the tool to run in.
Can a tool do more than the app's page?
No. A tool can only do what the page can already do. It gets no extra privileges and no access the app did not already have, which is what keeps the approval meaningful.
How do I make Ola choose my tool?
Give it a clear name and a plain description of what it does and what it changes. Ola uses the descriptor to decide when the tool fits. Narrow, well-described tools are chosen more reliably than broad ones with vague names.
WebMCP is a small idea with a useful shape: let the app publish what it can do, let Ola ask, and let the merchant approve. Build narrow tools, describe them honestly, and keep the API for the work that happens without a page. Read what Salesive apps are for the install model, and what Ola can actually do for the assistant's side. More guides are on the Salesive blog.