Claude for developers

How developers actually use Claude in 2026 - Claude Code in the terminal, the Agent SDK, IDE integrations, subagents, hooks, skills and plugins, plus what still goes wrong.

Most developers use Claude through Claude Code - an agentic coding tool that runs in the terminal, reads a codebase, edits files, runs commands and opens pull requests. It installs with one command, is included on every paid Claude plan but not on Free, and is extended through skills, subagents, hooks, plugins and MCP servers. The alternative routes are the Claude API and the Claude Agent SDK.

This page covers what a session actually looks like, how the extensibility model fits together, which IDEs are supported, and - in a table at the end - where it still falls down. Documentation moved: Claude Code docs now live at code.claude.com/docs, not under the old platform.claude.com path.

#How do you install Claude Code?

One command on macOS, Linux or WSL:

curl -fsSL https://claude.ai/install.sh | bash

On Windows, irm https://claude.ai/install.ps1 | iex in PowerShell. There are also packaged routes: brew install --cask claude-code, winget install Anthropic.ClaudeCode, npm install -g @anthropic-ai/claude-code (Node.js 22+), and signed apt, dnf and apk repositories. The distinction that catches people out is updates - native installs auto-update in the background; Homebrew, WinGet, apt, dnf and apk do not. Two release channels exist: latest by default, and stable, which trails by about a week and skips major regressions.

Requirements are modest: macOS 13.0+, Windows 10 1809+, Ubuntu 20.04+, Debian 10+ or Alpine 3.19+; 4 GB of RAM; x64 or ARM64; and a network connection, since all reasoning happens server-side. Release binaries are GPG-signed and notarised. Download routes for the rest of the Claude family are on the download page.

#What does an agentic coding session actually look like?

You run claude in a repository and describe an outcome, not a file. Claude explores the codebase itself - usually by dispatching a read-only Explore subagent rather than loading everything into context - proposes a plan, asks permission before each write or command, then edits, runs tests and iterates on the failures.

$ cd ~/src/billing-service
$ claude

> /init                       # writes a CLAUDE.md describing the repo
> /plan                       # switch to plan mode: research, no edits
> Refactor the retry logic in worker/ to use exponential backoff
  with jitter, and add tests for the 429 path.

  ● Explore(retry logic)                        18 files read
  ● Plan ready - 4 steps, touches 3 files       [accept / edit / reject]
  ● Edit worker/retry.py                        +52 -19
  ● Bash(pytest tests/test_retry.py)            2 failed
  ● Edit worker/retry.py                        +7 -3
  ● Bash(pytest tests/test_retry.py)            14 passed

> /diff                       # review before committing
> /code-review                # a second pass over your own change

Three mechanics do most of the work. CLAUDE.md is read at the start of every session - build commands, conventions, things not to touch - plus an auto-memory Claude maintains as it works. Plan mode separates research from execution, which stops a large refactor going sideways on step one. And permissions gate every tool call by default; widen them deliberately with /permissions.

Model choice matters here more than in chat. Opus 5 is the default for complex agentic coding, Sonnet 5 is the sensible workhorse for everyday edits, and /effort raises or lowers reasoning depth per session. Because usage is metered against your plan's shared allocation, dropping effort on routine work is the simplest way to make a Max plan last the day - see how the tiers compare.

#How is Claude Code extended?

Six mechanisms, layered. Most teams use two or three and ignore the rest, which is fine - the point is that they compose.

#Subagents versus agent teams

A subagent gets its own context window, does a job and returns a summary to the main conversation. That is the right shape for search, for a focused review, for anything where the intermediate reasoning is noise. Up to 20 run concurrently, nested three levels deep.

Agent teams are peers instead of subordinates: independent contexts, messaging each other under a structured protocol, no single orchestrator holding everything. They suit genuinely parallel work across a large codebase. They are off by default in cloud sessions and need an environment variable to enable, which is a fair signal of maturity.

#Plugins and marketplaces

A plugin is a self-contained directory with a .claude-plugin/plugin.json manifest that can carry skills/, commands/, agents/, hooks/hooks.json, .mcp.json, LSP config and background monitors. Anthropic runs two marketplaces: claude-plugins-official, curated and registered automatically on first launch, and claude-community, open to public submissions and added manually. Test locally with claude --plugin-dir ./my-plugin. Plugin skills are namespaced /plugin-name:skill-name, so collisions are not a concern.

MCP is the piece worth understanding properly, because it is the same protocol the chat apps and the API use - it was donated to the Linux Foundation's Agentic AI Foundation in December 2025 and is no longer Anthropic-specific. Server patterns and security considerations are on MCP and Claude agents, and scheduled or event-driven orchestration on the automation page. If you would rather wire workflows visually than write a server, Claude in n8n covers that route.

#Which IDEs and surfaces are supported?

The terminal CLI is the full-featured surface; everything else is a shell around the same engine.

#Claude Code on the web

Research preview At claude.ai/code, sessions run in isolated Anthropic-managed VMs with restricted network access and git credentials held outside the sandbox behind a proxy. Start one from the terminal with claude --cloud "task" and pull it back with claude --teleport. It is available to Pro, Max and Team users, and to Enterprise users on premium or Chat + Claude Code seats.

The constraints are real. GitHub is required for cloning and PR creation - other repos upload as a bundle under 100 MB but cannot push back. IP allowlisting breaks cloud-session authentication unless Anthropic's ranges are exempted, and organisations on zero data retention cannot use cloud sessions at all. There is no separate compute charge.

#Which plans include Claude Code, and what does it cost?

Claude Code is bundled into existing subscriptions with no separate fee. The Free plan does not include it - authentication requires Pro, Max, Team, Enterprise or a Console API key. On a subscription, usage draws from the same shared allocation as chat, Cowork, Design and Claude for Excel, governed by a rolling five-hour window plus a weekly cap, with Opus models on their own weekly counter. Anthropic publishes relative multipliers, not message counts.

Paying per token through a Console API key is the other option, and the right one for CI and shared automation. Some features are subscription-only regardless: Claude Code on the web, on mobile and in Slack, the desktop app, Routines, Code Review, Remote Control, computer use and voice dictation all require a claude.ai plan. Full plan mechanics are on the pricing page.

#What about the Agent SDK and building your own tools?

The Claude Agent SDK exposes the same agent loop, tools and context management that power Claude Code, as a library you run in your own process. It reads the same .claude/ and ~/.claude/ configuration, so skills, subagents, hooks and MCP servers carry across from your interactive setup.

It is available for Python and TypeScript only. To drive the same loop from Go, Rust or anything else, run the CLI as a subprocess with -p and --output-format json. Two distinct products sit either side of it: the client SDKs (seven languages, direct API access, you write the tool loop) and Managed Agents (Anthropic hosts the loop and sandbox behind a REST API, billed at token rates plus $0.08 per session-hour). Both are covered on the API page.

#Where does Claude Code still fall down?

An honest list, because the marketing version is not useful for planning.

The recurring theme is that the terminal CLI is the honest product and everything else is at some stage of catching up. If a feature matters to your workflow, check whether it exists on the surface you actually use before you plan around it.

#Frequently asked questions