Skip to content

Bananaflow MCP is live.

Read the docs

Claude Code MCP: Build Voice Agents by Chat

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Cover for "Claude Code MCP: Build Voice Agents by Chat": a white line diagram of a chat plugged into a voice agent: a chat bubble with three lines of text, its tail running on as a cable into a two-pin plug that meets a lit node, the agent, with a sound wave running out of it, under pills reading Chat, MCP and Voice agent, drawn over a soft, out-of-focus bellflower in muted aqua, its open bells blurred into veils, with a sky blue haze at one end and a dusty rose glow at the other.

Claude Code MCP is Claude Code's support for the Model Context Protocol, an open standard that lets an AI assistant use outside tools through servers. Bananaflow runs one of those servers, so Claude Code can create, edit, test and evaluate a voice agent while you describe it in plain words. This guide explains what each part is, gives the exact setup command, walks through one made-up build, and covers the habits that keep a live agent safe.

Here is the idea in a story. Farah runs admissions at a nursing college in Pune. Parents ring all day with the same questions about fees, seats and hostel rooms, and she wants an agent that answers in English and Marathi and offers a campus visit. She does not want to learn a builder first. She opens Claude Code, connects Bananaflow, and describes the call. Farah, the college and her choices are made up for this example.

What is Claude Code MCP?

Claude Code MCP is Claude Code's support for the Model Context Protocol, an open standard that lets Claude Code use outside tools and data through MCP servers. Anthropic's docs say "Claude Code can connect to hundreds of external tools and data sources through the Model Context Protocol (MCP), an open source standard for AI-tool integrations" (Claude Code MCP docs, read 15 November 2026).

People often ask what the difference is between Claude Code and MCP. Claude Code is the assistant you talk to. MCP is the plug it uses to reach a tool. Without a server on the other end, there is nothing to plug into.

What is an MCP server?

An MCP server is a program that offers tools and data to an AI application over the Model Context Protocol. The protocol's own introduction calls MCP "an open-source standard for connecting AI applications to external systems", and compares it to a USB-C port for AI applications (Model Context Protocol introduction, read 15 November 2026). For a plain-language explainer, read what an MCP server is.

The architecture page names three roles:

  • Host: the AI application you use, such as Claude Code or Claude Desktop.
  • Client: the part of the host that holds one connection to one server.
  • Server: the program that offers the tools. A remote server on the Streamable HTTP transport can serve many clients.

Do not confuse this with the tools an agent uses during a call. A voice agent can also call your own MCP servers while it talks to a customer. That is a different feature, covered on our integrations page. Here the MCP server is the way you build the agent, not something the caller touches.

Can an MCP server build a voice agent?

Yes: Bananaflow runs an MCP server, so Claude Code, Cursor, Codex or Claude Desktop can create, edit, test and evaluate Bananaflow voice agents and their tools. The server is at https://app.bananaflow.ai/api/v1/mcp/, uses the Streamable HTTP transport, and reads your API key from an X-API-Key header.

One chat, three results: Claude Code reaches Bananaflow's MCP server, which saves the agent as a draft, runs a text test and runs an eval.

The setup page lists every tool: Bananaflow's MCP setup guide. The product view is on the integrations page.

How do you add the Bananaflow MCP server to Claude Code?

Create an API key in Bananaflow, run claude mcp add with the HTTP transport and the key in the X-API-Key header, then check the result with claude mcp list. Claude Code's docs say HTTP servers are the recommended option for remote servers, and that the older SSE transport is deprecated.

  1. In Bananaflow, open the API keys page and create a key. The endpoint is also shown under Settings, on the MCP server tab.
  2. In a terminal, run the command below with your key in place of YOUR_API_KEY.
  3. Run claude mcp list. A working server shows as "Connected". The same list shows "Needs authentication" or "Failed to connect" when something is wrong.
claude mcp add --transport http bananaflow https://app.bananaflow.ai/api/v1/mcp/ \
  --header "X-API-Key: YOUR_API_KEY"

claude mcp list

If a client cannot send a custom header, the setup page says an Authorization: Bearer header works too.

One habit matters here. Claude Code can store a server at three scopes: local (this project, not shared), project (shared through version control in .mcp.json) and user (all your projects, not shared). Pick local or user for a server that holds a key. A project-scope file is committed with your code, so a key in it ends up in your repository.

What can you ask it to do?

You describe the call in plain words, and the assistant writes the agent, saves it as a draft and tests it by text before anything goes live. Here is Farah's session, in five prompts. The tool names are the ones the Build and Test groups of the setup page list.

  • Step: 1. Learn the rules

    What Farah types
    Read Bananaflow's voice prompting guide before you write anything
    Tools you may see
    get_voice_prompting_guide
  • Step: 2. Create the agent

    What Farah types
    Create an admissions enquiry agent for a nursing college in Pune. English and Marathi. It answers fee, seat and hostel questions and offers a campus visit
    Tools you may see
    create_workflow, get_workflow_code, save_workflow
  • Step: 3. Try it in text

    What Farah types
    Start a text chat as a parent asking about the BSc Nursing fee, and show me where the agent goes wrong
    Tools you may see
    start_text_chat, send_text_chat_message, end_text_chat
  • Step: 4. Lock in a test

    What Farah types
    Write an eval for the hostel question and run it
    Tools you may see
    create_eval, run_eval, get_eval_run
  • Step: 5. Go live

    What Farah types
    Publish it
    Tools you may see
    publish_workflow

What the assistant writes is a prompt, so the usual rules apply. Our voice agent prompt guide covers what a good one looks like, and our guide to building an AI calling agent shows the same job done by hand. The agent can answer in English and 10 Indian languages, including Marathi, as the languages page lists.

How do you test the agent before it takes calls?

Run a text chat and a saved eval against the draft, then test by voice in the browser, because text tests skip speech and turn timing. A text chat checks what the agent says. It cannot tell you how it sounds, whether it hears a name correctly, or how long it waits after a parent stops talking.

So use both:

  • Text chat: fast, and good for checking that the script follows the rules you gave.
  • Eval: a saved test you can run again after every edit, so a fix does not break something that worked.
  • Browser test by voice: the check for speech, accents and timing. Bananaflow lets you test by voice or text in the browser.

Our guide on how to test a voice agent lists the tricky cases to try before real callers do.

What stays under your control?

The assistant saves drafts. Your published agent keeps taking calls until you ask it to publish, and you can restore an older version. This is the most important safety fact in the post, so here are the rules in order:

  1. Edits go to the draft. The published agent keeps running as it was.
  2. The assistant publishes only when you ask. Read the draft first, then say "publish".
  3. Versions are kept. Two tools, list_workflow_versions and restore_workflow_version, let you see older versions and go back. The workflow builder shows the version history in the app.
  4. Tool edits go live at once. A change to a tool applies to every agent that uses it. Check which agents use a tool before you rename or change it.

Is it safe to give an AI assistant your API key?

Treat the key like a password: keep it out of Git and shared chats, connect only servers you trust, and review what the assistant plans to change. The setup page says to treat it like any other credential, and not to commit it to source control or paste it into shared chats.

Anthropic adds a warning of its own: "Verify you trust each server before connecting it. Servers that fetch external content can expose you to prompt injection risk." (Claude Code MCP docs, read 15 November 2026). Prompt injection means text the assistant reads, such as a web page, tries to give it new orders.

Two habits follow:

  • Read what the assistant plans to change before you approve it.
  • Keep the agent's own rules tight. Our guide to AI voice agent guardrails shows what to write into the agent itself.
Draft to live: edits land in the draft, and the published agent only changes when you ask the assistant to publish.

Build with AI in the app or Claude Code over MCP: which should you use?

Use the chat panel inside Bananaflow if you work in the browser, and the MCP server if you already work in a coding assistant. Both can write and edit an agent. They differ in where you sit.

  • Point: Where you work

    Build with AI in the app
    A chat panel beside the workflow
    Claude Code over MCP
    Your terminal or editor
  • Point: Good for

    Build with AI in the app
    People who live in the browser
    Claude Code over MCP
    Developers who already use a coding assistant
  • Point: What it does

    Build with AI in the app
    Writes and edits the workflow and creates tools as it goes
    Claude Code over MCP
    Creates, edits, tests and evaluates agents and their tools
  • Point: Setup

    Build with AI in the app
    None: it is in the app
    Claude Code over MCP
    One command and an API key

Either way the agent ends up in the same workflow builder, and you can open it there and edit by hand. The Bananaflow MCP server is separate from the feature that lets an agent use your own MCP servers during a call, which the integrations page explains.

How does Bananaflow fit?

Bananaflow is an AI voice agent platform for Indian businesses. Its MCP server lets Claude Code and other assistants build agents, while the app gives you the workflow builder, the Build with AI chat panel, browser testing, draft and publish with version history, and 100+ apps in one click. The setup guide is at docs.bananaflow.ai, and the product overview is on the integrations page.

What should you do next?

Try it on a small agent first. This week, create an API key, run the setup command, and ask the assistant for one short agent, such as a reception greeter. Read the draft, run a text chat, write one eval, and test by voice in the browser. Publish only when you would be happy for a customer to hear it.

FAQ

Frequently asked questions

  • What is the difference between Claude Code and MCP?

    Claude Code is the AI assistant you talk to. MCP, the Model Context Protocol, is the open standard that lets it reach outside tools, and an MCP server is the tool on the other end. Claude Code is the host, and it holds one client connection for each server it uses.

  • How do I add an MCP server to Claude Code?

    Run claude mcp add with the HTTP transport, a name, the server's address and any header it needs. For Bananaflow that is claude mcp add --transport http bananaflow https://app.bananaflow.ai/api/v1/mcp/ --header "X-API-Key: YOUR_API_KEY". Then run claude mcp list to see whether it shows as connected.

  • Does the Bananaflow MCP server work with Cursor or Codex?

    Yes. Bananaflow's MCP server lets an AI assistant such as Claude Code, Codex, Claude Desktop or Cursor create, edit, test and evaluate voice agents and their tools. If a client cannot send a custom header, the setup page says an Authorization Bearer header works too.

  • Can the assistant publish my agent without asking?

    It publishes only when you ask. Edits made through the MCP server save to the agent's draft, and your published agent keeps taking calls as it was until you ask for a publish. Older versions are kept, so you can restore one.

  • Is my published agent affected while I edit?

    No, the published agent is not changed by draft edits. The one exception is a tool: an edit to a tool goes live at once for every agent that uses it. Check which agents use a tool before you rename or change it.

  • Do I need to write code to build an agent this way?

    No. You describe the agent in plain words and the assistant writes it, saves it as a draft and tests it. You still read the draft, run a text chat and an eval, and test by voice in the browser before you ask it to publish.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot