Most people do not pick one MCP server: they add the servers of the services they already use. All ten here run at a hosted address, offer browser sign-in and give set-up steps for Claude, Cursor and VS Code. Six checks decide which ones to add, and how safely. Each rests on facts read on the vendors' own pages on 15 November 2026.
1. Start from the service your work lives in
An MCP server reaches one service, so the right choice depends on where your files, tickets or data already sit. A developer on GitHub adds the GitHub server. A team on Jira and Confluence adds Atlassian's. Web servers, such as Exa, Firecrawl and Bright Data, work on public web pages rather than inside an account.
The server rarely costs anything on its own. GitHub, Notion, Atlassian and Apify state that connecting is free or included in every plan. What you pay is the service behind it. A team of five on Jira Standard pays $42.40 a month for Jira, and nothing extra for the server.
2. Check that your AI app is on the vendor's list
Each vendor publishes set-up steps for the AI apps it supports. All ten cover Claude, Cursor and VS Code. Nine give steps for ChatGPT or for Codex, OpenAI's coding assistant; Bright Data covers only OpenAI's developer tools.
Follow the vendor's own steps rather than a generic guide. Several vendors, GitHub and Atlassian among them, offer one-click buttons that fill in the settings for you. If your app is missing from a vendor's list, test the server on a free plan before you rely on it.
3. Prefer a hosted server with browser sign-in
A hosted server runs at an address the vendor maintains, so there is nothing to install or update. All ten offer one, though Sentry's is still marked Beta. Browser sign-in, known as OAuth, opens a window where you log in and approve access, so you never copy a secret key into a settings file.
A key or token (a long password made for programs) still has a use. With it, a script or an automated job can connect without anyone at the keyboard. Notion keeps that option for its Business and Enterprise plans. If you want the server on your own machine, seven of the ten run there in full. Two more do it only in part: Supabase with fewer tools, and Notion with its unmaintained older server.
4. Decide what the assistant may change
Some servers only read; others can create, edit and delete. Exa and Bright Data change nothing in your accounts: they search and read the web. GitHub, Supabase, Notion, Atlassian and Sentry can change code, databases, pages, tickets or issues.
A server that can write needs a limit. GitHub has a read-only flag, Supabase can run every query as a read-only user, and Atlassian's admins can block edits. GitHub warns about prompt injection: hidden instructions in a file or web page that try to steer the assistant. Supabase warns of security risks in its docs. Start in read-only mode, then allow changes once you have watched the assistant work.
5. Count what each call costs
The price that matters is what each request uses up in the service. Exa charges per search, Bright Data per result, Firecrawl per page in credits and Context7 per call above the plan's allowance.
A worked example shows the gap. An assistant that runs 1,000 searches a month on Exa costs $4 at list price, which Exa's $10 of free monthly credits cover. On Context7 Pro, a developer who makes 3,000 calls a month goes 1,000 calls over the 2,000 a seat includes. That adds $5 to the $10 seat, so $15 a month. Check the usage page of each service after the first week.
6. Keep the list of tools short
Every tool a server exposes takes space in the assistant's working memory. GitHub notes that enabling only the needed toolsets helps the assistant choose the right tool. Seven of the ten let you pick tools or groups of tools.
Switch on only what the task needs, and add tools when a new task calls for them. On GitHub, toolsets do this by group. With Apify, Exa, Bright Data or Sentry, a parameter in the server address does the same job.