The best MCP servers for business: sales, marketing and revenue (2026)
A verified directory of MCP servers worth connecting for business work: Stripe, HubSpot, Google Analytics, PostHog, Notion, Linear and more, grouped by job, with what each one actually exposes.
By Pedro Campos

Most "best MCP servers" lists are written for developers: filesystem access, browser automation, database clients. Useful, but not what you need if you run a company. This list is the other one: the servers that make an AI assistant useful for revenue, sales, marketing and operations, chosen the way you would choose a vendor rather than a repo to star.
Three rules governed what got in:
- Official first. A server run by the company whose data it touches, with OAuth you can revoke, beats a community wrapper you have to hand an API key to.
- Verified. Every remote endpoint listed below was probed before being written down. No dead links, no vaporware announcements.
- Business jobs only. If its main audience is someone writing code, it is not here.
One disclosure up front: Kometrics, in the revenue section, is our product. It is listed under the same rules as everyone else, and the rest of the list stands on its own.
What is an MCP server?
MCP (Model Context Protocol) is the open standard that lets AI assistants use outside tools. An MCP server exposes a set of typed tools; clients like Claude, ChatGPT, Cursor and Codex connect and can then call them mid-conversation. For you, the practical meaning is: instead of pasting exports into a chat, the assistant fetches live data itself, and instead of you clicking through a dashboard, it acts on your behalf. Most servers below connect with a URL and an OAuth approval, no code involved.
Revenue and billing
The money section has a split worth understanding: some servers operate on money (refunds, invoices, payment links) and some explain it (MRR, churn, LTV). You will likely want one of each, and we wrote up why the operational ones cannot compute your metrics separately.
Stripe
Official, by StripeThe Stripe API behind four meta-tools, plus docs search: refunds, payment links, invoices, disputes. It writes as well as reads, so keep tool confirmation on.
https://mcp.stripe.comDocsKometrics
Official, by KometricsComputed subscription metrics instead of raw billing objects: MRR, churn, LTV, cohorts, and the movement ledger behind them. Read only by construction.
https://app.kometrics.com/mcpConnectPayPal
Official, by PayPalInvoices, payments, disputes and catalog across a PayPal business account, in chat.
https://mcp.paypal.com/mcpDocsSquare
Official, by SquarePayments, customers and catalog against the Square API, for businesses that sell on Square.
https://mcp.squareup.com/mcpDocsQuickBooks and Xero, the accounting half of the money stack, only have community servers as of this writing: usable, but read their code before handing over a key.
Sales and CRM
HubSpot
Official, by HubSpotContacts, companies, deals and tickets. Ask your pipeline questions, or update records without opening the CRM.
https://mcp.hubspot.comDocsZapier MCP
Official, by ZapierOne server that proxies thousands of apps (Salesforce, Pipedrive, Mailchimp and the rest) as MCP tools. The pragmatic route to any CRM that has no server of its own.
https://mcp.zapier.comDocsSalesforce, Pipedrive and most other CRMs do not ship their own server yet, which is exactly the gap Zapier's MCP fills: one connection, and every app Zapier reaches becomes a tool your assistant can call.
Marketing and analytics
Google Analytics
Official, by GoogleTraffic, acquisition and behavior reports from your GA4 properties. Runs locally (Python) rather than as a hosted endpoint.
PostHog
Official, by PostHogProduct analytics, feature flags, experiments and error tracking, queryable in chat.
https://mcp.posthog.com/mcpDocsSemrush
Official, by SemrushKeyword volumes, domain rankings, backlinks and competitive data. The keyword research behind this article ran through it.
https://mcp.semrush.com/v1/mcpDocsAd platforms are the notable hole. Meta Ads and Google Ads have no official MCP servers yet, only community ones, and an ads account token is a credential you should be slowest of all to hand to unaudited code. Until official servers land, keep ads reporting on exports.
Ops and internal tools
Notion
Official, by NotionPages and databases: meeting notes, wikis and the content calendar, readable and writable.
https://mcp.notion.com/mcpDocsLinear
Official, by LinearIssues, projects and cycles. Triage, status updates and sprint questions from chat.
https://mcp.linear.app/mcpDocsAtlassian
Official, by AtlassianJira issues and Confluence pages over Atlassian’s remote server.
https://mcp.atlassian.com/v1/sseDocsHow to actually use these
A few things we learned running a metrics MCP server in production, from both sides of the connection:
- Two servers beat ten. Every connected server spends part of the model's attention. The pattern that works is one operational server for the thing you do most, plus one analytics server for the questions you ask most, and everything else added when a task needs it.
- Know which connectors can write. Stripe's server can refund and cancel; Notion's can edit pages. Keep the client's tool confirmation on for those, and prefer read-only servers (or read-only keys) everywhere you only ask questions.
- OAuth is a feature, not a formality. Official servers let you see and revoke sessions from the vendor's dashboard. A community server holding a raw API key gives you no such lever, which is the real difference between the two badges above.
- Ask "why", not just "what". The point of giving an assistant your pipeline or your MRR is not reading numbers aloud, it is the follow-up: why did this move, which customers drove it, what happens if the trend holds. Servers that expose computed, per-entity data make those follow-ups answerable; servers that expose raw lists make the model guess.
FAQ
Are MCP servers free?
The servers themselves are, almost universally: they are a feature of the underlying product, not a separate purchase. You pay for the product behind them (the CRM seat, the analytics plan) exactly as before. Kometrics, for example, is free under $1,000 MRR, connector included.
Do MCP servers work with ChatGPT?
Yes. ChatGPT added MCP support through connectors in Developer mode, and the Codex CLI takes any server as a config entry. Claude, Claude Code and Cursor connect to the same URLs, which is the point of MCP being a standard: one server, every assistant.
Are MCP servers safe to connect?
Treat them like team members with credentials. Official servers with OAuth and revocable sessions are the safe default; community servers deserve a code read first. The main risk to manage is write access combined with prompt injection, so keep tool confirmation on for anything that can move money or edit data.
What is the difference between an MCP server and a regular integration?
An integration syncs data between two products on a schedule. An MCP server gives a language model tools it can decide to call mid-conversation, in whatever combination the question requires. The first automates a pipeline you designed; the second lets the assistant improvise one per question.
Know your revenue. Trust the metrics.
Connect Stripe and get every SaaS metric computed from your real billing history. Free under $1,000 MRR.
Start freeKeep reading

How to connect Stripe to Claude (and actually ask about your revenue)
Two ways to connect Stripe to Claude over MCP: the official Stripe server for payment operations, and Kometrics for MRR, churn and LTV your agent can reason over. Setup for Claude, Claude Code and Codex.

The Stripe MCP server: what it does, how to set it up, and what it can't tell you
How the Stripe MCP server works, how to add it to Claude, Claude Code, Cursor and Codex, the full tool list, and the revenue questions it cannot answer, plus how to fix that.

A ChartMogul alternative built for agents, not just dashboards
Kometrics was built for the AI era: your revenue inside Claude, ChatGPT, Claude Code and Codex over MCP, an AI that reads your live book, and a ledger every answer can be traced back to. How it compares to ChartMogul, and how to switch.