The Claude Coding Tools Stack: What to Install in 2026
"Coding with Claude" in 2026 isn't one tool — it's a stack with four layers: the agent, the editor surface, the connectors, and the expertise layer on top. Most developers install layer one, skim layer two, and never discover that layers three and four are where the compounding happens. Here's the whole map.
Layer 1: The agent — Claude Code
The foundation. Claude Code is Anthropic's agentic coding tool: it reads your repo, edits files, runs commands and tests, and iterates until the task is done — under a permission system you control. It runs in the terminal, inside VS Code and JetBrains IDEs, on the web, and headlessly in CI.
- Install:
npm install -g @anthropic-ai/claude-code, thenclaudein any project. - Best for: everything with a definition of done — features, bugs, refactors, migrations, code review, test coverage.
- First move: run
/initto generate a CLAUDE.md. Our getting-started guide covers the rest.
Layer 2: The editor surfaces
IDE extensions (VS Code, JetBrains)
Claude Code's IDE integrations put the agent next to your editor: inline diff review, file context awareness, and the same sessions you'd have in the terminal. If you live in an IDE, start here rather than the raw terminal.
AI-native editors (Cursor and friends)
Editors like Cursor blend completion with an agent and let you choose Claude as the model. A reasonable choice if you want one app — but note the division of labor: editor agents are strongest at editor-scoped work, while Claude Code's loop (run, verify, iterate) and extensibility go deeper for repo-scale tasks.
GitHub Copilot with Claude models
Copilot can run Claude models as its backend — Claude-quality completions inside the tool your company may already license. For how completion and delegation actually divide the work, see Claude vs Copilot.
Claude on the web and GitHub app
Claude Code's web/cloud sessions handle work when you're away from your machine, and the GitHub integration reviews PRs and responds to issues — useful as the team-wide surface that doesn't require every member to change tools.
Layer 3: The connectors — MCP servers
The Model Context Protocol gives Claude tools beyond your filesystem: query the database, drive a browser for end-to-end verification, read Sentry errors, search your issue tracker. For development work, the highest-value connectors are browser automation (Playwright-based servers), database access (Postgres et al.), and your error tracker. Full breakdown with setup notes: our MCP server guide.
Layer 4: The expertise — skills and commands
The layer most developers never install. Claude Code is extensible with skills (folders that teach it repeatable capabilities) and slash commands (reusable prompts invoked by name). This is how a generic agent becomes your agent — one that reviews code against your standards, writes tests your way, and audits security like a specialist.
- Free, from us: the SKILL.md validator (lint the skills you write) and the website security audit skill.
- The full library: the ClaudeThings Engineering Kit — 58 agents, 61 skills, and 159 commands covering review, testing, security, refactoring, and shipping. One command to install.
What to install first: the minimal path
- Claude Code (terminal or IDE extension) + a CLAUDE.md in your main repo.
- One MCP server you'll actually use — browser automation or your database.
- One skills library — write your own starting from our validator, or install a kit.
- Stop. Add more only when a real task demands it; tool-collecting is procrastination with better branding.