Integrations
Integrations that let your voice agent act during the call
Give your agent tools it uses while the caller waits: your own APIs, tested before they go live, your MCP servers, and apps such as HubSpot, Gmail and Google Calendar that connect in one click.

What are AI voice agent integrations?
AI voice agent integrations are the tools a voice agent calls during a live phone call to read or write data in other systems, such as looking up an order or booking a slot.
In Bananaflow a tool is made once and attached to the steps that need it, so a greeting step cannot touch your billing system.
New to voice agents? Read how a voice agent call works, one turn at a time.
Building agents from a coding assistant? Read how to set up Bananaflow's MCP server in Claude Code.
Wondering what an agent can check on a live call? See how an inbound agent looks up an order or a payment.
New to MCP? Read what an MCP server is.
For the open standard behind MCP servers, see the Model Context Protocol site.
How do you connect a voice agent to your systems?
You create a tool once, attach it to the steps of the call where it is needed, and tell the agent in plain words when to use it. Setup takes five steps.
Store your key once
Save it as a credential: an API key, a Bearer token, Basic auth or a custom header. A tool then picks it by name.
Create a tool
Choose an HTTP API, an MCP server or an app. Describe each parameter in plain words, so the agent knows what to send.
Test an HTTP tool
Run it with sample values and check the reply and the time it took, before any caller reaches it.
Attach it to the steps that need it
Pick the steps of the call that may use the tool, and say in the step's prompt when to use it.
Make a browser test call, then publish
Talk to the agent in your browser, watch it use the tool, and publish when it behaves.
Not sure how your systems would connect?
What can the agent connect to?
Any HTTPS API, any MCP server that speaks Streamable HTTP, and apps such as HubSpot, Gmail, Google Calendar, Slack and Stripe that you connect with their own sign-in. Four things make that useful.
Call any API you already run
Point a tool at your own HTTPS endpoint and the agent can use it mid-call.
check_slots(pincode: "500081", date: "2026-11-04")
- GET, POST, PUT, PATCH or DELETE to your own endpoint, with your key kept in a stored credential.
- The agent fills the parameters you describe. Bananaflow fills the rest, like the caller's number, so the agent never has to ask.
- An error comes back to the agent as an error, so it never tells a caller that a failed booking went through.
Test a tool before a caller does
Run an HTTP tool with sample values and see exactly what comes back.
200 OK, 3,180 ms
- See the status code, the reply and the headers that were sent. It warns you before it runs a method that changes data.
- See how many milliseconds it took, so you know when the caller needs a hold line.
- Set that hold line as text or as your own recording. It plays while the request runs, on HTTP tools and app tools.
Bring your own MCP server
If your team already built an MCP server, the agent can use its tools.
get_report_status
- Paste the server's URL and pick a credential. Refresh lists the tools it offers.
- Allow only the tools you want, then choose again for each step of the call.
- If the server is down, the call carries on without its tools.
Connect 100+ apps, like HubSpot or Gmail, in one click
Sign in on the app's own screen, then pick one action for the agent to run. Over 100 apps connect this way.
- Click Connect and sign in on the app's own consent screen. There is no key to copy.
- Pick one action, like creating a HubSpot contact or a Google Calendar event.
- Fill the fixed fields once, so the agent is only asked for what the caller says.
Use LeadSquared, Zoho, Salesforce or another system with an API? Reach it through an HTTP tool that calls its API with your key. Check your system's API docs for the endpoints you need.
The test panel is for HTTP tools. For an MCP server, Refresh lists its tools. For an app tool, try it in a browser test call. Sign-in for apps follows OAuth 2.0, the open standard.
How does a call center use integrations?
A lab's booking team lets the agent check slots in its own system and log each booking in HubSpot, so no one retypes the call. This example is made up: the names and numbers are illustrative, not a customer.

Monday, 10:00 am
The slot check gets tested
A diagnostic lab chain in Hyderabad takes 900 home-sample booking calls a day in Telugu, Hindi and English. Developer Karthik Rao builds check_slots as an HTTP tool against the lab's booking API and runs a test: 200 OK in 3,180 ms. Three seconds of silence is too long, so he adds a hold line: "Ek second, main slot dekh rahi hoon."
11:30 am
The team's MCP server goes in
He adds the MCP server his team already built for the lab's website chat. It offers eight tools. He allows only one, get_report_status, on the "Report enquiry" step.
12:15 pm
HubSpot connects in a click
Ops lead Lakshmi Reddy clicks Connect HubSpot and signs in. Karthik adds a "Create contact" action and fills the caller's number from the call, so the agent never asks for it.
Tuesday
Test calls, then publish
They make browser test calls in all three languages, fix two prompts and publish.
Support teams use the same tools to look up orders and update tickets. See how support calls get resolved with tools.
How does Bananaflow work with your CRM?
Before the call it can fetch the caller's record, during the call it can read and update it, and after the call it can send the outcome to it. Three moments, three tools.
Before the call
A pre-call data fetch asks your system who is calling and fills variables like {{customer_name}}, so the greeting can use the caller's name before the first word.
During the call
Template variables and preset parameters send the caller's number or the run ID for you, so the agent is never asked for something Bananaflow already has.
After the call
Webhooks send the call's outcome and recording link to your CRM, and call events can also go to BigQuery.
See how to fetch the caller's record before the greeting in the workflow builder, and how to send QA results with the call outcome by webhook.
Workflow builder
Attach tools to each step of the call and fetch the caller's record first.
AI customer support calls
Look up orders and update tickets while the customer is on the line.
AI inbound call handling
Check an order or a payment on calls that come in.
Appointment confirmation calls
Book and reschedule through your calendar.
Calling campaigns
Run a list and pass each row's values into the call as variables.
Call QA
Score each call you choose and send the result with the outcome by webhook.
FAQ
Integration questions, answered
What are AI voice agent integrations?
AI voice agent integrations are the tools a voice agent calls during a live phone call to read or write data in other systems, such as looking up an order or booking a slot. In Bananaflow they are HTTP API tools, your own MCP servers and apps that connect with a sign-in.
Does Bananaflow work with LeadSquared, Zoho or Salesforce?
Yes, through an HTTP API tool that calls the CRM's own API with your key. Check your CRM's API docs for the endpoints you need, then test the tool in Bananaflow before you attach it.
Which apps connect in one click?
Over 100 apps connect in one click with their own sign-in (OAuth), including Gmail, Stripe, HubSpot, Google Calendar and Slack. An app that has no one-click sign-in can still be reached through an HTTP tool that calls its API with your key.
Can the agent use our own MCP server?
Yes. Add its URL as an MCP Server tool, and the agent can call the tools you allow while it is on the call. The server must support Streamable HTTP.
Can Claude Code or Cursor build agents over MCP?
Yes. Bananaflow also runs its own MCP server, so a coding assistant such as Claude Code or Cursor can list your agents and edit them for you.
What does the caller hear while a tool runs?
Your own hold line, typed or recorded, plays while the request runs. You set it on each HTTP tool and each app tool.
What happens if our API is slow or down?
HTTP tools stop waiting after 5 seconds by default and app tools after 10, and the agent is told the call failed, so it can say so instead of guessing. You can change the HTTP timeout.
Where are our API keys kept?
In stored credentials that a tool refers to by name. The agent sees only the tool's name, description and parameters, never the key.
Can the agent know who is calling before it speaks?
Yes. A pre-call data fetch can ask your system for the caller's record and pass the fields into the greeting as variables.
Bring us one process. We'll pilot it on your real calls, with your team as the fallback.
Plan a pilot