The best MCP servers
An MCP server gives your AI (Claude Code, Claude Desktop, Cursor, VS Code…) access to a service: documentation, browser, GitHub, files, memory, web search.
7 scored mcp. Last checked on 13 September 2026.
The three to remember
The most installed MCP, and for good reason: it eliminates hallucinated APIs.
- · Utility 9/10, best of the selection
- · Nothing to flag on security
- · Official publisher
Sufficient for discovering persistent memory, too simple for serious use.
- · One-click install
- · Free, no limits
- · Your data stays on your machine
The right compromise for anyone who wants transparent agent memory under their own control.
- · Free, no limits
- · Your data stays on your machine
- · Nothing to flag on security
| Tool | Price | Score | Sec. |
|---|---|---|---|
| Context7MCP server that injects up-to-date, version-specific library documentation into the agent's context. | Freemium | 9.4 | Security: nothing to flag |
| Brave Search MCPRisingIndependent web search for the agent via the Brave API. | Freemium | 8.5 | Security: nothing to flag |
| Basic MemoryRisingAgent memory stored as local Markdown files, readable in Obsidian. | Free | 8.4 | Security: nothing to flag |
| Playwright MCPRisingGives the agent a controllable browser: read pages, click, fill forms, test. | Free | 8.0 | Security: read before installing |
| GitHub MCPRisingOfficial GitHub MCP server: issues, PRs, code, and Actions accessible to the agent. | Free | 8.0 | Security: read before installing |
| Filesystem MCPRisingReference MCP server for reading and writing files within authorized folders. | Free | 7.8 | Security: read before installing |
| Memory MCP (officiel)RisingReference knowledge graph memory server, stored in a local file. | Free | 7.3 | Security: read before installing |
Tool : standalone product, used directly · Plugin : installs into a host application · Skill : instructions that teach the ai a method · MCP : connector giving the ai access to a service
What to know
An MCP server runs on your machine or at a vendor and exposes “tools” the agent can call. Every install adds attack surface: read the security dot and requested permissions before wiring a community server, and run mcp-scan after each addition.