# What is an MCP server? A guide for marketing and RevOps teams Canonical URL: https://www.martechdevs.com/blog/what-is-an-mcp-server By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo) Data Engineer and Co-founder, martechdevs Published: 2025-09-22 Last updated: 2026-10-06 An MCP server gives an AI application a defined way to retrieve information or perform actions in another system. For a marketing team, that could mean reading HubSpot deals, querying GA4 or checking Google Ads performance. What it can access depends on the server’s tools, the connected account’s permissions and the controls around each action. HubSpot holds the deal. Stripe holds the payment. GA4 holds the visit. Connecting an AI assistant to one of them lets it work with that system’s information. Answering a question across all three still requires a reliable way to connect the records. That distinction matters when someone asks which campaigns brought customers who stayed. The assistant needs customer identity, payment history and an agreed definition of retention before it can produce an answer worth taking into the meeting. [Preparing data for AI access](https://www.martechdevs.com/ai-ready-data) starts with those definitions. ## What does MCP mean in plain English? MCP stands for Model Context Protocol. It defines how an AI application discovers and calls capabilities supplied by another program. The [MCP architecture documentation](https://modelcontextprotocol.io/docs/learn/architecture) describes a host, a client and a server. The host is the AI application you use. Its client communicates with a server. The server exposes specific capabilities, such as retrieving a report or looking up a record. Servers can run on your computer or be hosted remotely. You don’t need to learn the protocol messages to assess an integration. You need to know who runs the server, which account it uses and what its available tools actually do. Picture an assistant investigating a slow sales week. A useful sequence is: 1. Read the agreed pipeline definition. 2. Retrieve this week’s deals and the comparison period. 3. Check whether the source data is complete. 4. Calculate the difference. 5. Explain the result with links to the underlying records. The connection makes the retrieval possible. The definition, checks and calculation make the answer useful. ## How does an MCP request reach your marketing data? The AI application chooses an available tool and supplies its inputs. The server checks the request, calls the underlying system and returns a result. The application then uses that result in its response. ![An AI application selects a tool, the MCP server checks access, and HubSpot returns permitted deals with their scope for the AI to explain.](https://www.martechdevs.com/assets/blog/what-is-an-mcp-server.svg) The important boundary sits between the requested action and the source system. An assistant asking for every customer record should still be limited by the authenticated user’s access and the server’s rules. A useful answer should carry enough detail to reproduce it: the account, date range, filters, metric definition and retrieval time. “Pipeline fell 12%” leaves too much hidden. “New deal value created this week, excluding renewals and using the sales team’s reporting time zone” gives the reader something they can check. ## What can marketing and RevOps teams use it for? Start with recurring work where the inputs and expected output are already clear. That makes errors easier to spot and savings easier to measure. ### Investigate a report An assistant can retrieve a campaign report, compare periods and suggest which change deserves investigation. Ask it to separate observed changes from explanations that still need evidence. A drop in paid traffic is an observation. A claim that the new landing page caused it needs additional evidence. The assistant should show that distinction in its answer. ### Prepare an account or pipeline review An assistant can gather permitted CRM records and prepare a briefing. Define the fields it should use and the point at which it must stop. For example, a sales review might need deal stage, value, next activity and close date. It probably doesn’t need every email body, support conversation or attached contract. ### Draft operational changes A server with write tools may let an assistant propose or apply updates. HubSpot’s [Spring 2026 developer announcement](https://developers.hubspot.com/changelog/spring-2026-spotlight) describes write access for specified CRM objects through its remote MCP server. Treat each action separately. Adding a task, changing a deal owner and enrolling contacts in communication workflows have different consequences. An integration that can read contacts hasn’t automatically earned permission to change them. For an initial workflow, have the assistant prepare a change list showing the record, current value, proposed value and reason. The person reviewing it should be able to understand the effect without opening another conversation. ## How is MCP different from an API, connector or AI agent? An API is an interface software uses to interact with another system. An MCP server can wrap selected API operations so an AI application can discover and call them. The original API’s permissions, limits and data model still matter. A connector is a product term. It may use MCP underneath, or it may use another integration method. Read its documented capabilities before assuming that two connectors with the same vendor name expose the same data. An AI agent is the system deciding which steps to take towards a task. MCP can supply some of the tools it uses. The agent also needs instructions, controls and a way to decide whether the task succeeded. A skill supplies reusable working instructions. For example, “review weekly pipeline” could specify the date logic, exception checks and report format. The MCP connection supplies access to the records. Our guide to [context engineering for marketing](https://www.martechdevs.com/blog/context-engineering-marketing) explains how to keep those instructions usable. These parts solve different problems. When an answer is wrong, identify the failing part before changing the whole setup: - **No records returned:** check account access, tool inputs and filters. - **Correct records, wrong total:** check the calculation and definition. - **Correct total, wrong explanation:** check whether the assistant inferred a cause without evidence. - **Unexpected updates:** check tool permissions and the approval process. ## Which marketing tools have an MCP server? Availability and capabilities change by vendor. Use the vendor’s documentation as the starting point, then inspect the tools exposed in your own connection. Google maintains an [Analytics MCP repository](https://github.com/googleanalytics/google-analytics-mcp) and a separate [Google Ads MCP repository](https://github.com/googleads/google-ads-mcp). HubSpot documents its own remote server. These are separate integrations with separate access requirements. The practical comparison is broader than whether a server exists: - Does the vendor maintain it, or does another company run it? - Is it hosted remotely, or do you run software locally? - Which reports, objects and actions are available? - Which user or service account authorises access? - Where does retrieved data go after the call? - Who handles updates when the source API changes? Use the [marketing MCP server guide](https://www.martechdevs.com/blog/mcp-servers-for-marketing) to compare connection options. Start with a system that contains most of the evidence for one useful task. ## Does an MCP connection make AI answers accurate? It gives the assistant access to evidence. Accuracy still depends on which evidence it retrieves and how it interprets it. Consider the question “What was customer acquisition cost last month?” The answer changes with the definition: - Advertising spend divided by new paying customers. - Advertising plus agency costs divided by new paying customers. - All sales and marketing costs divided by new customers. Each calculation can be arithmetically correct. Only one may match the number finance expects. Write the agreed definition into a maintained reference or governed reporting model. Include the treatment of refunds, currencies, late payments and customer reactivations. The assistant should flag missing inputs before returning a confident number. Cross-system questions also need a reliable match between records. A company domain can help identify an account, but subsidiaries, shared domains and personal email addresses complicate it. If the join is uncertain, return the unmatched records separately rather than quietly dropping them. ## What does a sensible first project look like? Use one recurring question, one source and a known answer. Add complexity after the team can reproduce the result. **Illustrative example:** a 12-person SaaS company prepares a weekly HubSpot pipeline review. The operations lead spends 90 minutes gathering deals, checking missing next steps and formatting the report. The pilot asks an assistant to retrieve the same deal set, group it by owner and highlight missing next activities. It excludes free-text emails and makes no CRM changes. The operations lead compares every output against the existing review for four cycles. Suppose gathering and formatting falls to 20 minutes, while review still takes 25 minutes. The workflow saves 45 minutes each week. At an assumed internal cost of $60 an hour, that is $45 a week, or $180 across four weeks. Those are modelled inputs, not a claim about typical MCP savings. That estimate also sets a spending limit. If additional software, maintenance and troubleshooting exceed the saved effort, expand the business case only when another measurable benefit exists. Faster reporting alone doesn’t justify every integration. Record the pilot’s accuracy as well as its time. A report that saves an hour and misstates the largest deal needs more work. ## Is an MCP server safe for company data? Safety depends on the complete connection: the server, the authenticated account, the AI application and the tools available in the same session. “Read only” limits changes to the source, but retrieved data still leaves that source. Check the AI provider’s data handling, retention and workspace controls for the product and plan you actually use. Also distinguish a tool’s description from enforced permissions. MCP maintainers explain that [tool annotations are hints](https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/). A tool labelled read only still needs trustworthy implementation and appropriate access controls. A malicious instruction can also arrive inside content the assistant reads. For example, a CRM note could tell it to send a customer list elsewhere. Treat retrieved business content as evidence to inspect. It should not gain authority over the user’s task or access policy. The [MCP security checklist](https://www.martechdevs.com/blog/mcp-security-checklist) covers permission tests, data handling and revocation in more detail. ## What should you check before connecting one? Copy this checklist into the pilot ticket: ## Checklist - [ ] Name the recurring task and the person who will review its output. - [ ] Save a known answer with its filters, date range and metric definition. - [ ] Confirm the server’s publisher through the vendor’s own documentation. - [ ] List the exposed read and write tools. - [ ] Use the smallest account access that can complete the task. - [ ] Identify which fields can leave the source system. - [ ] Set a clear approval step for changes with business consequences. - [ ] Test missing data, empty results and denied access. - [ ] Check the answer against the underlying records. - [ ] Record the time saved, maintenance effort and failures. - [ ] Test how to disconnect the integration and revoke its credentials. At Martechdevs, we build the data access, definitions and working instructions behind repeatable AI workflows. Start with the [free martech AI audit](https://www.martechdevs.com/martech-ai-audit), or explore an [AI assistant built around your approved workflows](https://www.martechdevs.com/custom-martech-ai-agent). ## Frequently asked questions ### What is an MCP server URL? It is the address a compatible client uses to reach a remotely hosted server. A local server may instead start from a command on your computer. Take connection details from the publisher’s setup documentation, because an ordinary website address may not be the MCP endpoint. ### Do I need a developer to use one? A supported hosted connector can be straightforward to authorise. Running a local server, setting up cloud credentials or building a custom connection usually needs someone comfortable with the relevant tools. Cross-system reporting may need data modelling even when connecting the server is easy. ### Can an MCP server change my CRM? It can when its exposed tools and underlying permissions allow that action. Inspect the actual tool list and test with limited access before using it on important records. An instruction to avoid changes is useful guidance alongside enforced restrictions. ### Does MCP replace my data warehouse? A connection does not create shared customer identity, payment history or agreed metrics across your systems. A warehouse can provide those joined records and definitions when the use case needs them. An [internal MCP server](https://www.martechdevs.com/blog/internal-mcp-server) can then expose focused questions over that prepared data. ### What should I connect first? Connect the source that answers a frequent, narrow question with a result someone can check. A weekly pipeline review or a fixed campaign report is a useful candidate. Add additional systems when the task requires their evidence. ## Sources 1. Model Context Protocol, [Architecture overview](https://modelcontextprotocol.io/docs/learn/architecture), current documentation, accessed October 2026. 2. HubSpot, [Spring 2026 Spotlight](https://developers.hubspot.com/changelog/spring-2026-spotlight), 2026. 3. Google, [Google Analytics MCP](https://github.com/googleanalytics/google-analytics-mcp), repository, accessed October 2026. 4. Google, [Google Ads MCP](https://github.com/googleads/google-ads-mcp), repository, accessed October 2026. 5. Model Context Protocol maintainers, [Tool annotations as risk vocabulary](https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/), 2026.