MCP lets an AI assistant like Claude read from and write to your tools, instead of working only with what you paste into the chat. For a team that prospects by phone, that means pulling up a contact's history, preparing a call list or creating follow-ups in one sentence. Flunter publishes an MCP server built on its API. This article explains what MCP is in plain terms, what our server can do today, how to install it and where its limits are.
In short
An MCP server for cold calling connects your AI assistant to your calling tool. With Flunter's MCP server, you ask Claude, in plain language, to find a contact by name or phone (by company or email within a given campaign) and its call history, prepare a list, create a campaign ready to dial or schedule follow-ups. The assistant goes through the Flunter API with your key and cannot do anything that key does not allow. The calls themselves still happen in Flunter, with a rep on the line.
What is MCP, and why does it matter for a sales team?
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems: databases, software, files. Anthropic released it in November 2024, and the official protocol website compares it to a USB-C port: one standard plug, so nobody has to build a different connection for every tool and assistant pair. Claude, ChatGPT, Visual Studio Code and Cursor are among the clients listed.
In practice, an MCP server describes a list of tools to the assistant ("search a contact", "create a task"...) with their parameters. When you ask a question, the assistant picks the right tool, calls it, reads the answer and sums it up for you.
For a sales team, the benefit is very down to earth. Out in the field, teams never work in a single tool: a CRM (more and more often a homemade one on Airtable), spreadsheets, messaging apps, dashboards. A good part of the day goes into copying and pasting between them. MCP does not replace your CRM or your parallel dialer. It removes some of that back and forth, letting the assistant handle the lookups and data entry while you focus on calls.
The tools in Flunter's MCP server
The server is called @flunter/mcp (npm package). Each tool maps to an endpoint of the Flunter API. Here are the 15 tools available today, grouped by use.
Use | Tools | What it does | Access |
Find a contact | search_contacts, get_contact | Searches by name or phone (by company or email within a given campaign), then returns the full record with its notes and tasks | Read |
Update a contact | update_contact | Changes the fields you provide (job title, email, company, "do not call"...); custom fields and status cannot be changed | Write |
Manage campaigns | list_campaigns, get_campaign, create_campaign | Lists campaigns, opens one with its contacts, or creates a new one and imports a contact list into it | Read and write |
Read calls | list_calls, get_call | Lists calls with filters (contact, campaign, dates, duration, status, outcome, inbound or outbound) and reads one call | Read |
Understand the setup | get_contact_fields, get_call_outcomes | Returns the organization's contact fields and call outcomes with their groups (success, follow up, refusal, unreachable, bad contact) | Read |
Organize follow-ups | list_tasks, get_task, create_task, update_task, delete_task | Lists, creates, updates or deletes tasks linked to a contact (call, email, other) | Read, write, delete |
For calls, the API returns each call's duration, status and outcome and, when available, the transcript, the AI summary and the link to the recording. That is what makes the assistant useful: it can reread what was said before preparing a callback for you.
Most write tools are described to the assistant as tools to use only on explicit request, and depending on the MCP client, you may be asked to confirm before they run.
Real use cases, day to day
Get the context back before a callback
"Find Claire Martin at Dupont Industrie and summarize our last three calls." The assistant searches the contact, lists the calls, reads the summaries and gives you the essentials: what was said, the objection from the last call, the next step you promised.
Prepare a call list
You have a trade show list, a spreadsheet export or contacts read from another tool connected to the assistant. You ask: "Create a campaign called Lyon trade show follow-up with these 40 contacts." The assistant first checks your organization's required fields, creates the campaign, then flags the rejected contacts (invalid number, missing field) while the others are imported.
Sort things out after a session
"List today's calls marked follow up and create a callback task for each one, on the date agreed in the summary." This is where the time saved shows the most.
A worked example
Take a rep who ends a two-hour session with, say, 12 conversations, 5 of which need a follow-up. Creating 5 tasks by hand, with the right contact, the right date and a note on what was said, easily takes ten minutes. With MCP, it is one request, a quick review and a confirmation. These figures are an example to make things concrete, not a measurement.
Update a record
"Mark this contact as do not call, they asked me to." The contact is then excluded from calls. Or: "Their new title is Head of Sales."
Install Flunter's MCP server
You need two things: a Flunter API key, generated in the app (Settings, then API, then "Generate a key"), and Node.js version 20 or later on your computer. Nothing to download by hand: the client runs the package on the fly with npx.
Client | How to add it |
Claude Code | In a terminal, run: claude mcp add flunter-mcp --env FLUNTER_API_KEY=your_key -- npx -y @flunter/mcp |
Claude Desktop | In the claude_desktop_config.json file, add a "flunter-mcp" entry under "mcpServers" with the command "npx", the arguments "-y" and "@flunter/mcp", and the environment variable FLUNTER_API_KEY set to your key. Then restart Claude Desktop. |
Other MCP clients | Same idea: the command npx -y @flunter/mcp and the FLUNTER_API_KEY variable. |
Two optional variables exist: FLUNTER_API_URL (defaults to https://api.flunter.com) and FLUNTER_API_TIMEOUT_MS (maximum time for a request, 30 seconds by default). The full list of tools, with their parameters and access level, is published in the Flunter MCP documentation.
Once installed, test with a read-only question such as "List my five latest campaigns". If the answer comes back, everything works.
The API behind MCP, and the homemade CRM angle
The MCP server is only a thin layer on top of the Flunter REST API. Anything it does, a developer can do directly:
Address: https://api.flunter.com, with endpoints under /v1.
Authentication: the API key, sent in the x-api-key header.
Limits: per user, across all keys, a burst of 10 requests per second and 2,500 requests per hour. Beyond that, the API returns "429 Too Many Requests".
Documentation: api.flunter.com/v1/doc for the API, api.flunter.com/v1/mcp/doc for the MCP server.
Price: API access is included in every plan, from the Starter plan at €89 excl. VAT per license per month.
When we launched the API in April, I put it this way: when Flunter becomes the calling engine, it has to fit cleanly into the team's ecosystem. If your CRM is one of the native integrations (HubSpot, Salesforce, Pipedrive, Close, Boond Manager, Odoo), the sync is already done. If you work on a homemade CRM, an Airtable base or custom dashboards, the API is the right path: create campaigns from your own lists, fetch every call with its transcript and recording, feed your cold calling KPIs.
MCP comes on top: the API is for automations that run on their own, MCP is for one-off requests you make in plain language.
Security and limits to know
What the server can do, and nothing more
The MCP server has exactly the rights of the API key it is given. It sees the campaigns, calls and tasks the key owner sees, nothing else. A few simple rules:
One key per use: one for MCP, one for each automation. You can have up to 10 active keys.
A short lifetime when possible: a key expires after one week, one month, six months or one year, as you choose (one year by default).
The key is shown only once: store it in a password manager, never paste it into a shared document.
Delete a key as soon as it is no longer used, or at the slightest doubt.
Review before confirming a campaign creation, a contact update or a task deletion.
What MCP does not do today
No statistics or user list: the API offers them, but the MCP server does not expose them yet. For dashboards, use the API or the app's call reporting.
No calls launched from the assistant: MCP prepares and organizes, conversations happen in Flunter.
Contact search in two steps: the API has no direct contact search, so the tool scans at most 10 matching campaigns. On a large base, name the campaign.
What about ChatGPT?
According to OpenAI's developer documentation, checked on October 1, 2026, ChatGPT connects to remote MCP servers, added by web address in developer mode. Flunter's server runs locally on your computer, launched by npx. So today it is designed for Claude Code, Claude Desktop and clients that run local MCP servers. Codex, OpenAI's development tool, is one of them: its MCP documentation, checked on October 1, 2026, says it supports servers started locally by a command. Each assistant's offering changes fast: check what your client supports when you read this.
Frequently asked questions
What is an MCP server for a CRM?
It is a small program that describes to an AI assistant the actions available in your CRM or calling tool (search, read, create, update) and runs them through its API. The assistant understands your request in plain language and picks the action.
Do I need to code to use Flunter's MCP?
No. You generate an API key in Flunter, have Node.js installed and add one command or a few lines of configuration. After that, everything happens in plain language.
Can MCP make calls for me?
No. The server reads and organizes your contacts, campaigns, calls and tasks. Calls happen in Flunter, and a rep leads the conversation.
What is the difference between Flunter's API and MCP?
The API is the technical interface, used by your developers or automation tools. MCP is a layer on top of it so an AI assistant can use it. MCP covers contacts, campaigns, calls and tasks; the API also offers statistics and the user list.
Is my data safe with MCP?
The server only has the rights of your API key, which expires and which you can delete at any time. Keep in mind that API responses are read by the AI assistant you use: choose it according to your company's rules.
Can I use Flunter's MCP with ChatGPT?
Not directly today: according to OpenAI's documentation checked on October 1, 2026, ChatGPT connects to remote MCP servers, while Flunter's runs locally. On the OpenAI side, Codex does accept local MCP servers. Claude Code and Claude Desktop remain the simplest clients to start with.
What you can do this week
Generate an API key dedicated to MCP, with a one-month lifetime to start.
Install the server in Claude Code or Claude Desktop, then test a read-only question.
Pick one repetitive task to hand over to the assistant, such as end-of-session follow-ups, and do it with the assistant for a week.
Write down the time saved and any mistakes before extending it to other uses or other reps.
If you have a homemade CRM, show the API documentation to whoever maintains it.
For official, up-to-date information about Flunter (features, pricing, API), see our information for AI assistants page. And if you do not use Flunter yet, plans and the trial are on the pricing page.