Integrations
Use apps as agent tools
Switch on app actions and the built-in web tools in an agent's Toolbox, so it can look things up or act while it answers. How write actions are confirmed, limits and traces.
Guide3 min readUpdated 31 October 2026
An agent can call a connected app's actions while it answers, the way it calls MCP tools: look up an order in Shopify, find a HubSpot contact, open a Linear issue. Two built-in tools need no app at all: Read web pages and Search the web.
Before you start
- An agent. See Create an agent.
- The apps it should use, connected. See Connect an app.
Steps
- 1
Open the agent's Toolbox
Open the agent and go to its Build tab. The Toolbox ("What it can use while it answers.") lists Built in tools and, under Connected apps, every action of every app you've connected, one group per connection. Point at an action to see what it does.
- 2
Switch on what it may use
Turn on the actions it needs. Each switch saves straight away, and the count at the bottom shows how many are on, out of 30 (app actions and built-in tools together; MCP tools don't count). Connect more apps takes you to Integrations and back.
- 3
Confirm actions that change data
Actions that post, send, create or update carry a Changes data badge. Turning one on asks first: "Anyone who chats with this agent can make it … with your … account." Choose Turn on only if that's what you want.
- 4
Tell the agent when to use them
Say in its instructions when each tool is for, for example "Look the order up in Shopify before you answer about it." See Write a system prompt.
- 5
Try it
Ask a question that needs the tool in the Build tab's test chat or the Playground. The reply shows each tool call with its input and output.
You're done when
A test question makes the agent call the tool, and its trace shows a tool.execute span for it.
Built-in tools
| Tool | What it does | Needs |
|---|---|---|
| Read web pages | Opens a public link someone gives it, or one it finds, and reads the page's text | Nothing |
| Search the web | Looks things up on the web with Tavily when its knowledge doesn't have the answer | A connected Tavily account. With no Tavily key to use, the agent simply isn't given this tool |
How agents call app actions
- The model sees each action as a tool named
<app>_<action>, such asslack_post_message, with the action's description and fields. - A call has 15 seconds. A failure goes back to the model as an error message, so it can tell the person; the reply doesn't stop.
- The model reads at most the first 2,500 characters of a tool's answer, as with MCP tools, and has up to 8 tool rounds per reply.
- Tool calls cost no extra credits: the reply is 1 credit however many tools it calls.
- Each call is a
tool.executespan in the trace withtool.sourceset tointegration. See Trace reference. - An agent with app tools never answers from the answer cache.
- Experiments give the agent only the built-in tools, so running a dataset never posts or creates anything.
Public agents and actions that change data
Anyone who can chat with the agent (on your widget or a chat link) can lead it to use its tools, write actions included, and a page or search result it reads can carry instructions of its own. Give a public agent read-only actions unless posting or creating records at a visitor's request is what you want.
Tools picked when the agent was created
On New agent, a Give it tools chip switches on every action of that app, including ones that change data, without asking, and drafting can add tools too. Check the Toolbox before you put a new agent on your website.
Apps or MCP?
Use an app when SentientOne has it: you connect with a token and there's nothing to host. Use an MCP server for your own systems and anything not listed. An agent can have both; the Toolbox shows its MCP servers in one row with Manage, which opens the agent's Tools tab.
Still stuck?
Email us with what you tried and any request id. A person replies.