14 November 2026 · 4 min read
The best MCP servers for coding agents in 2026
Five servers I would add to Claude Code, Codex or Cursor, and what each one fixes.
IThe short answer
Start with three: the GitHub MCP server for issues and pull requests, Context7 for current library docs, and Playwright MCP to test the web pages the agent builds. Add Chrome DevTools MCP when you debug performance or network problems. Add offstage when agents must run graphical apps without taking over your screen.
Every server adds tools to the agent's context. Turn on only the toolsets you use.
IICompared
| What it fixes | Runs | License | Add it | |
|---|---|---|---|---|
| GitHub MCP server | The agent can read repos and manage issues, PRs and Actions | Remote (hosted by GitHub) or local | MIT | Remote URL https://api.githubcopilot.com/mcp/ |
| Context7 | Outdated or invented APIs: it fetches current, version-specific docs | Remote or local | MIT (server); backend is private | npx ctx7 setup |
| Playwright MCP | The agent can open and test pages through accessibility snapshots | Local | Apache-2.0 | npx @playwright/mcp@latest |
| Chrome DevTools MCP | Performance traces, network and console debugging in Chrome | Local | Apache-2.0 | npx -y chrome-devtools-mcp@latest |
| offstage (mine) | Graphical apps run on a separate screen, not on your desktop | Local, on Linux or Windows | MIT | mrwick1/offstage |
From each project's README, checked on 14 November 2026.
IIIGitHub MCP server
GitHub's official server lets an agent read code and manage issues, pull requests and workflows. The remote server needs no install and signs in with OAuth or a personal access token. The local server runs as a Docker image. The default toolsets are context, repos, issues, pull_requests and users. --read-only removes the write tools, which is a good default for a new agent.
IVContext7
Context7 from Upstash feeds current library documentation into the agent's context. Add "use context7" to a prompt, or name a version such as Next.js 16 to get the matching docs. It has two tools: resolve-library-id and query-docs. A free API key gives higher rate limits. The docs are community-contributed, so check what the agent copies.
VPlaywright MCP and Chrome DevTools MCP
Playwright MCP lets the agent click and type by element reference from an accessibility snapshot, so it needs no vision model. It drives Chrome, Firefox, WebKit and Edge. Chrome DevTools MCP controls a live Chrome and adds performance traces, network requests and source-mapped console errors. Its usage statistics are on by default.
I compare both with my own from-scratch browser in Playwright MCP vs Chrome DevTools MCP vs bare.
VIoffstage
offstage gives each agent its own screen, so a browser or app under test never takes your focus or cursor. I maintain the fork and wrote most of its current code, so read this as the author's view. The agent reads the screen as text first, and each action returns a verdict such as CONFIRMED or NO_OP. It works on Linux, where a screen is an in-memory compositor, and on Windows in a smaller form, where the agent drives a headless Bare browser. It does not isolate apps from the file system or network.
VIIQuestions
Which MCP servers should I add to Claude Code?
A good starting set is the GitHub MCP server for repositories, issues and pull requests, Context7 for current library documentation, and Playwright MCP for testing web pages. Add Chrome DevTools MCP for performance and network debugging.
What does Context7 do?
Context7 fetches current, version-specific library documentation and code examples into a coding agent's context, which reduces outdated or invented API calls. Its MCP tools are resolve-library-id and query-docs.
Does GitHub have an official MCP server?
Yes. github/github-mcp-server is MIT-licensed. GitHub hosts a remote version at https://api.githubcopilot.com/mcp/, and a local version runs as the Docker image ghcr.io/github/github-mcp-server.
How do I stop too many MCP tools from filling the context?
Enable only the toolsets you need. The GitHub MCP server, for example, accepts --toolsets to choose groups and --read-only to remove write tools.