How do MCP servers work with coding agents?

A Model Context Protocol (MCP) server exposes tools, such as 'create a pull request' or 'fetch library docs', over a standard JSON-RPC protocol. The coding agent (Claude Code, Codex, Cursor and others) is the MCP client: it lists the server's tools, sends their names and schemas to the model, and when the model chooses a tool, the client calls the server and returns the result to the model. Local servers run as a subprocess over stdio; remote servers run over Streamable HTTP with OAuth. The current specification is MCP 2026-07-28.

Which MCP servers are most useful for coding agents?

As of October 2026, five earn their place in most setups: GitHub's official MCP server (issues, pull requests, code search; remote at api.githubcopilot.com/mcp/), Playwright MCP from Microsoft (drive a real browser to test the app you just changed), Context7 from Upstash (current library documentation instead of stale training data), your issue tracker or error monitor (Linear, Jira, Sentry), and a database server scoped to read-only access. The reference Git and Filesystem servers matter mainly for agents that lack built-in file and git tools.

How do I add an MCP server to Claude Code, Codex or Cursor?

Claude Code: claude mcp add --transport http for a remote server, or claude mcp add -- for a local one; project servers go in .mcp.json. Codex: codex mcp add, or a [mcp_servers. ] table in ~/.codex/config.toml or a trusted project's .codex/config.toml. Cursor: install from the Cursor Marketplace or add the server to mcp.json; Cursor supports stdio, SSE and Streamable HTTP.

How many MCP servers should a coding agent have?

As few as the task needs, usually three to five. Every connected server adds its tool names and schemas to the model's context on each request, which costs tokens and makes wrong tool choices more likely. Enable toolsets selectively (GitHub's server supports a --toolsets flag) and prefer read-only access where the agent does not need to write.

Quick Answer

How MCP Servers Work With Coding Agents, and Which to Use

Published: October 19, 2026

The short answer

An MCP server is a small program that offers tools to an AI agent over a standard protocol; your coding agent is the client that lists those tools, hands them to the model, and runs the ones the model picks. For coding, the servers that pay for themselves are GitHub’s official server, Playwright MCP, Context7, your issue tracker or error monitor, and a read-only database connection. Keep the total small: every server’s tool list rides along in every request. Setup commands and versions below were checked October 19, 2026.

How it works, step by step

Connect. The agent starts a local server as a subprocess (stdio transport) or connects to a remote URL (Streamable HTTP, usually with OAuth). Since the 2026-07-28 revision, MCP is stateless: the client declares its protocol version on every request instead of a one-time handshake.
Discover. The agent calls tools/list and receives each tool’s name, description and JSON schema.
Offer. Those definitions go into the model’s context alongside your prompt.
Call. When the model decides to use a tool, the agent sends tools/call with arguments, usually after asking your permission.
Return. The server runs the action — a GitHub API call, a browser click, a SQL query — and the result goes back to the model for its next step.

That is why MCP matters for coding agents: the same GitHub or Playwright server works in Claude Code, Codex, Cursor, VS Code and others, instead of each tool shipping its own integration.

Adding a server in the three main agents

Agent	Remote server	Local server	Where config lives
Claude Code	`claude mcp add --transport http notion https://mcp.notion.com/mcp`	`claude mcp add playwright npx @playwright/mcp@latest`	`.mcp.json` (project), `~/.claude.json` (user)
Codex (CLI, IDE, app)	`codex mcp add <name> --url <url>` (OAuth flags supported)	`codex mcp add` with a command	`~/.codex/config.toml` or `.codex/config.toml` (trusted projects), `[mcp_servers.<name>]`
Cursor	Marketplace one-click install with OAuth, or a URL in `mcp.json`	Command in `mcp.json`	`mcp.json`; transports stdio, SSE, Streamable HTTP

Claude Code accepts streamable-http as an alias for http in JSON config and falls back to SSE automatically on v2.1.265 and later.

The most useful servers for coding agents

Server	Maintainer	What it gives the agent	Version / status (Oct 2026)	Use when
GitHub MCP Server	GitHub	Issues, PRs, code search, Actions; remote endpoint `api.githubcopilot.com/mcp/`	v2.0.2 (Oct 8, 2026)	Agent should open PRs, read issues, check CI
Playwright MCP	Microsoft	Drives a real browser via accessibility snapshots	v0.0.83 (Sep 28, 2026)	Verifying UI changes, reproducing bugs end to end
Context7	Upstash	Current, version-specific library docs (`mcp.context7.com/mcp`, or `npx ctx7 setup`)	ctx7 0.8.0 (Oct 16, 2026)	Fast-moving frameworks the model knows badly
Issue tracker / error monitor	Linear, Atlassian, Sentry	Ticket context, stack traces	Vendor-hosted remote servers	”Fix the bug in ENG-4521” workflows
Database (read-only)	Your DB vendor or community	Schema and query access	Varies	Debugging data issues; never give write access by default
Reference servers (Git, Filesystem, Fetch, Memory)	MCP project	Basic building blocks	Active in `modelcontextprotocol/servers`	Agents without built-in file or git tools; learning MCP

Claude Code, Codex and Cursor already read files, edit code and run git themselves, so the reference Filesystem and Git servers add little there. Several older reference servers (GitHub, GitLab, Brave Search) were moved to an archive repository — use the vendors’ official servers instead.

Rules that keep MCP useful and safe

Fewer tools, better choices. Enable only the toolsets you need; GitHub’s server has a --toolsets flag for this.
Trust before install. Anthropic’s docs warn that servers fetching external content expose you to prompt injection. Prefer official, vendor-maintained servers.
Least privilege. Fine-grained, short-lived tokens; read-only database users; no production write access. See short-lived credentials for AI agents.
Govern at scale. Teams with many agents put MCP traffic behind a gateway — see vendors extending API gateways into agent governance.

More server picks: best MCP servers, updated registry guide. Building one: how to build an MCP server. How MCP compares with plugins and skills: agent plugins vs MCP vs agent skills.

Last verified: October 19, 2026. Versions are each project’s latest GitHub release on that date.

The short answer

How it works, step by step

Adding a server in the three main agents

The most useful servers for coding agents

Rules that keep MCP useful and safe

Sources