Best MCP Servers for Coding Agents (2026)

Steps

1. 01

   Understand the tool-bloat rule before connecting anything

   Every MCP server you connect adds tools to the agent's context window. Coding agents in 2026 work best when they have 2–3 focused servers active at once. Connecting more than 5–7 servers causes tool bloat: the model spends token budget choosing between tools rather than writing code, and performance degrades noticeably. The practical pattern is: add the servers you actually need for the task at hand, and remove or disable others. All three clients let you list and remove servers at any time.

2. 02

   Context7 — live library documentation (fixes outdated-API hallucination)

   Context7 fetches up-to-date documentation for the libraries in your project and injects it into the agent's context. Without it, agents hallucinate deprecated APIs because their training data is months or years stale. This is the single highest-value server for day-to-day coding work. Add to Claude Code (stdio transport, confirm the exact package name on the Context7 docs):

   claude mcp add context7 -- npx -y @upstash/context7-mcp

3. 03

   GitHub MCP — repos, issues, PRs, CI, and code search

   The GitHub MCP server exposes your repositories, issues, pull requests, CI status, and code search as agent tools. Your agent can open a PR, check a failing CI run, or search across repos without you switching context. Note: the npm package `@modelcontextprotocol/server-github` is deprecated (last published 2025, npm marks it as no longer supported). GitHub's officially supported server is `github-mcp-server`, available as a hosted remote HTTP endpoint. Add it to Claude Code using the HTTP transport (confirm the exact URL on GitHub's MCP documentation):

   claude mcp add --transport http github https://api.githubcopilot.com/mcp --header "Authorization: Bearer your_pat_here"

4. 04

   Playwright MCP — browser automation and UI testing

   The Playwright MCP server gives the agent a real browser it can navigate, click, screenshot, and assert against. Use it when you need the agent to test a UI change, scrape a page that requires JavaScript, or run end-to-end checks as part of a coding task. Add to Claude Code (stdio, confirm the exact package name on the Playwright MCP docs):

   claude mcp add playwright -- npx -y @playwright/mcp

5. 05

   Exa — semantic web search (the most-used search MCP in 2026)

   Exa runs neural/semantic search over the live web. Your agent can research an error message, find an example implementation, or check whether a library has a known issue — without leaving the coding session. It requires an Exa API key. Add to Claude Code (stdio with API key env var — confirm the exact package name on exa.ai docs):

   claude mcp add exa -e EXA_API_KEY=your_exa_key_here -- npx -y exa-mcp-server

6. 06

   Desktop Commander — terminal and filesystem access

   Desktop Commander exposes shell execution, file read/write, and process management as MCP tools. It is useful for agents that need to run build scripts, inspect logs, or manipulate files as part of a coding workflow — particularly in OpenCode or Codex where you want those operations to go through the structured MCP layer rather than raw tool calls. Add to Claude Code (stdio — confirm the exact package name on the Desktop Commander docs):

   claude mcp add desktop-commander -- npx -y @wonderwhy-er/desktop-commander

7. 07

   Add MCP servers in OpenCode

   OpenCode uses the same MCP protocol but its own CLI surface. Use `opencode mcp add` to register a server interactively, `opencode mcp list` (alias: `ls`) to see what is connected, and `opencode mcp auth` for servers that use OAuth. List registered servers and their status:

   opencode mcp list

8. 08

   Add and inspect MCP servers in Codex

   Codex stores server configuration in ~/.codex/config.toml. Use the `codex mcp` subcommands to manage servers from the CLI, or edit the TOML file directly following the format in the Codex docs. If something is wrong, `codex doctor` will diagnose install, config, auth, and runtime health in one pass. List servers registered in Codex:

   codex mcp list

9. 09

   Verify, get details, and remove servers

   All three clients expose get and remove commands. Use them to inspect a server's registered config or clean up servers you no longer need (keeping active count at 2–3). Claude Code:

   claude mcp get <name>
   claude mcp remove <name>
   claude mcp list

10. 10

    Handle the approval flow in Claude Code

    When a server is added to a project-scoped .mcp.json file it appears as pending-approval until you explicitly approve it in the Claude Code session. If tools from a newly added server are not showing up, check whether the server is still pending. Run `claude mcp reset-project-choices` to reset all approved/rejected decisions for the project's .mcp.json servers and go through the approval flow again.

    claude mcp reset-project-choices

Troubleshooting

Newly added server tools are not showing up in the agent session (Claude Code)

The server may be sitting in pending-approval state. Project-scoped servers added via .mcp.json require explicit approval. Run `claude mcp reset-project-choices` and restart the session to go through the approval flow. Also run `claude mcp list` to confirm the server is registered and check `claude mcp get <name>` for its config.

npx: command not found when Claude Code tries to start an MCP server

Claude Code launches MCP server processes using your shell's PATH. If npx is not on PATH for the shell that Claude Code inherits (common after nvm installs or when launching from an app rather than a terminal), add the node/npm bin directory to your PATH in ~/.zshrc or ~/.bashrc and relaunch the terminal before running claude.

OpenCode shows an MCP server as disconnected after `opencode mcp add`

Run `opencode mcp list` to see the server's status. For servers that require OAuth, use `opencode mcp auth <name>` to authenticate and `opencode mcp debug <name>` to inspect the OAuth connection. For stdio servers, verify the command path is correct and the binary is executable.

Codex does not pick up a server I added to ~/.codex/config.toml

Run `codex doctor` first — it diagnoses install, config, auth, and runtime health in one pass and will surface config parse errors. Then run `codex mcp list` to confirm Codex sees the server. TOML syntax errors (missing quotes, wrong section headers) are the most common cause of silent failures.

Agent performance gets sluggish or it ignores tools after connecting several servers

This is tool bloat. Connecting more than 5–7 MCP servers floods the model's context with tool definitions it does not use for the current task. Remove servers you are not actively using with `claude mcp remove <name>` (or the equivalent in your client) and keep the active count at 2–3 for the task at hand.