Model Context Protocol (MCP) servers let an AI assistant read from and write to the tools you already use. Instead of copying data into a chat window, you connect the tool once and the assistant can query it directly. This guide lists the MCP servers that are actually worth connecting in 2026, grouped by what you are trying to get done.
We have organised this by job rather than by vendor, because the useful question is not “which MCP servers exist” but “which one removes the manual step I keep repeating”.
What an MCP server does, in one paragraph
MCP is an open standard that describes how an AI application talks to an external system. An MCP server is a small program that sits in front of a tool such as a CRM, a database or a file store, and exposes that tool as a set of actions the assistant can call. The assistant does not need custom code for each integration. It reads the list of available actions and picks the right one. If you want the longer explanation, read our guide on what an MCP server is.
How we chose the servers on this list
- It exists and runs today. No proposals, no waitlists.
- It does something you cannot do faster by hand. A server that wraps a single web form is not worth the setup.
- It is maintained. Either an official vendor build or a community project with recent commits.
- The permissions are legible. You can tell what it will read and what it can change before you approve it.
Disclosure: Scalelist publishes this guide and Scalelist offers its own MCP server. It appears in the sales and marketing section below, described using the same fields as every other entry.
Best MCP servers for business, sales and marketing
This is the category where MCP has moved fastest, because sales and marketing work is mostly moving records between systems. These servers remove the copy-paste layer.
Scalelist MCP server
Turns a plain-English request into a list of contacts with verified work emails and mobile numbers. You ask for the kind of company and role you want and the server returns the matching people, already enriched, so the assistant can hand them straight to your sequencer or CRM.
Best for: building a prospect list without leaving the assistant. Watch for: it consumes credits per enriched record, so scope the request before you run it.
HubSpot MCP server
Official server from HubSpot. Reads and writes contacts, companies, deals and notes, and can run association lookups. In practice the most valuable use is not creating records but reading them: asking questions across the CRM that would otherwise need a saved report.
Best for: CRM hygiene and ad-hoc pipeline questions. Watch for: write scopes are broad, so start read-only.
Salesforce MCP server
Exposes SOQL queries and record operations. Because Salesforce schemas are heavily customised, the assistant needs to discover your object and field names first, which the server supports. Once it has done that, it handles reporting questions well.
Best for: querying a customised Salesforce org without writing SOQL by hand. Watch for: permission sets apply, so the assistant sees exactly what the connected user sees.
Slack MCP server
Reads channels and threads, searches history, and sends messages. The read side is the useful side: pulling the last two weeks of a customer channel into a summary is a genuine time saver.
Best for: turning channel history into a written summary or a task list. Watch for: give it a bot token scoped to specific channels rather than a user token.
Gmail and Google Calendar MCP servers
Search threads, draft replies, list and create events, and find open slots. Drafting rather than sending is the safe default, and both servers support it.
Best for: inbox triage and scheduling. Watch for: keep send permission off until you trust the output.
Stripe MCP server
Reads customers, subscriptions, invoices and payment failures. Useful for revenue questions that do not justify building a dashboard, such as listing every failed payment in the last week with the customer attached.
Best for: billing questions and dunning lists. Watch for: use a restricted API key.
Notion MCP server
Searches pages and databases, creates pages, and updates properties. It is the most common place teams put the output of everything else on this list.
Best for: writing results somewhere a human will actually find them. Watch for: database queries are rate-limited on lower plans.
Linear MCP server
Lists, creates and updates issues. Pairs well with the Slack and GitHub servers: read a bug report from a channel, check the code, file the issue.
Best for: closing the loop between a conversation and a tracked task. Watch for: it will happily create duplicates, so ask it to search first.
Best MCP servers for developers
GitHub MCP server
Reads repositories, files, issues and pull requests, and can open branches and PRs. The read path alone justifies it: pointing an assistant at a repo it can actually browse produces far better answers than pasting files in.
Filesystem MCP server
Gives the assistant scoped access to a directory on your machine. Everything else is optional; this one changes how the assistant works day to day.
Postgres MCP server
Runs read-only SQL against a database and inspects the schema. Start it with a read-only role. There is no reason to give a language model write access to production data.
Playwright MCP server
Drives a real browser: navigate, click, fill forms, screenshot. Useful for checking that a page renders correctly and for reaching systems that have no API.
Sentry MCP server
Pulls issues and stack traces. Combined with the GitHub server, the assistant can read the error and then read the code that produced it.
Best MCP servers for data and research
Fetch and web-browsing servers
Retrieve a URL and return clean text rather than raw HTML. The single most common addition to any setup, because most questions eventually need a page the assistant has not seen.
Apify MCP server
Runs pre-built scrapers and returns structured datasets. Better than a generic fetch when the target is a site with pagination or login.
Google Drive MCP server
Searches and reads documents and spreadsheets. Read-only is usually enough.
Memory and sequential-thinking servers
These do not connect to anything external. One stores facts across sessions, the other structures multi-step reasoning. Low cost to add, noticeable effect on longer tasks.
Quick comparison
| Server | Main job | Start read-only? | Official build |
|---|---|---|---|
| Scalelist | Find and enrich contacts | N/A | Yes |
| HubSpot | CRM records and reporting | Yes | Yes |
| Salesforce | Query a custom org | Yes | Yes |
| Slack | Read and summarise channels | Yes | Community |
| Gmail / Calendar | Triage and scheduling | Yes | Community |
| Stripe | Billing and failed payments | Yes | Yes |
| Notion | Write results somewhere findable | No | Yes |
| Linear | Create and update issues | No | Yes |
| GitHub | Repos, issues, pull requests | Yes | Yes |
| Filesystem | Local file access | Yes | Reference |
| Postgres | Schema and SQL | Yes | Reference |
| Playwright | Browser automation | N/A | Yes |
| Apify | Structured scraping | N/A | Yes |
How to connect one
- Pick the client. Any MCP-compatible assistant will do; the configuration file location differs but the format does not.
- Add the server to the configuration, with its command and any environment variables it needs.
- Put credentials in environment variables, never in the file you might share.
- Restart the client and confirm the server appears in the connected tools list.
- Ask the assistant to list the actions it can now call. If the list is empty, the server failed to start.
- Run one harmless read before you grant anything write access.
Common mistakes
- Connecting ten servers on day one. Each one adds actions the assistant has to choose between. Add them as you need them.
- Granting write access immediately. Read-only first, every time.
- Using a personal token where a service token exists. A personal token inherits everything you can see.
- Assuming a failure is the protocol. Most connection problems are a wrong path or a missing environment variable.
Frequently asked questions
Do I need to write code to use an MCP server?
No. Most are installed by adding a few lines to a configuration file. Writing your own server requires code; using an existing one does not.
Are MCP servers safe?
An MCP server has exactly the permissions of the credential you give it. Scope the credential narrowly and start read-only, and the risk is the same as any other integration.
Can I run one on my own infrastructure?
Yes. Many run locally as a process on your machine. Hosted servers are also available for tools that need a persistent connection.
How many should I run at once?
Three to six covers most workflows. Beyond that, action lists get long and the assistant picks the wrong tool more often.
