Giving your agent tools

A tool is one specific thing your agent can do in an outside system, like get an order's status or cancel an order. Tools arrive from integrations, so connecting Shopify is what gives your agent Shopify's tools, and without an integration your agent can only talk.

That distinction runs through this whole screen, so it is worth getting straight first.

Integrations and tools are not the same thing

You connect integrations. You turn on tools.

• An integration is the system: Shopify, Amphora, Appstle, SendCloud, Rever. You connect it once, and it appears under Connected with a green dot.
• A tool is one action inside it. The Amphora integration alone brings Get Order Status & Tracking, Get Return Status, Hold / Release Order, Cancel Order, Update Shipping Address and more.

So one integration typically gives you eight to twenty tools, and you decide about each one separately. Turning on an integration does not turn on everything inside it.

Which is why the answer to "can my agent cancel an order?" is never just "yes, we use Shopify". The integration is connected, and Cancel Order is either off, gated, or open. Those are three very different products from your customer's point of view.

Where to find them

Open your agent, then Tools. The left column splits into Connected and Available, with Manage integrations at the bottom. Each integration has a single switch, Enable for this agent, and its tools listed underneath.

With more than one agent, this is per agent. Connecting Shopify to the workspace does not hand it to every agent.

Read tools and write tools

The tool list is split in two, because the two halves carry completely different risk.

• Read tools look things up. Get Order Status & Tracking, Get Order Line Items, Get Return Status, Get Delivery Proof / Label.
• Write tools change something in your business. Cancel Order, Update Shipping Address, Edit Order Line Items, Create / Approve Return.

The four modes

Every tool sits in one of four modes, and this is where you set how much rope your agent gets:

1. Off. The tool does not exist as far as your agent is concerned. This is also how you block one.
2. Verify by email or phone. Before using it, your agent checks the details against the email or phone on the order. Treat this as a speed bump rather than proof of identity: it confirms the person knows the details, not that they are the customer.
3. Verify with a one-time code. Your agent sends a passcode and waits for it. Slower, and much harder to fool.
4. Open. Your agent uses it freely, without checking who it is talking to.

Which mode to actually choose

Match the mode to what the tool exposes, not to how much you trust the AI.

Open is right for anything that reveals nothing personal. Searching your product catalogue, checking a discount code: there is no harm in a stranger getting an answer, and gating it just adds friction to a question you want answered instantly.

Verify by email or phone is the sensible default for reading someone's own data. Order status is the big one, and it is the difference between "where is my order" being answered in three seconds or in three hours.

A one-time code belongs on anything that moves money or changes a delivery. Refunds, cancellations, address changes. The friction is the point.

Off is a legitimate choice, but be clear about what it does. Your agent cannot use the tool and will not ask you to do it either, so a customer requesting that action will be told it is not something you can do here. If you want the request to reach you instead, leave the tool on and use a tool that creates a human task, or set your agent's instructions to hand those conversations over.

Bringing your own

Two entries in the list are not pre-built integrations:

• Custom HTTP tool points your agent at any API you have, with a name, a method, the authentication and the inputs it expects.
• Custom MCP connection connects a remote MCP server by its URL, and its tools appear alongside the rest.

Both are on or off only. The four verification modes apply to Konvo's own integrations, where we know what each action does.