Claude Code for VSCode

Original Article ↗ Hacker News Discussion ↗

Installation & Extension Behavior

  • Claude Code’s VSCode extension now appears on the Marketplace but was already being auto-installed when launching Claude Code from the VSCode terminal; some see this as convenient, others as invasive.
  • The extension adds: selection-to-context, diff viewing in the IDE, tab awareness (open files, selected text), keyboard shortcuts, and use of LSP error info instead of separate CLI checks.
  • For some users it previously behaved a bit buggy (e.g., uninstalling itself), with hope that the Marketplace release stabilizes things.

Claude Code vs Cursor / Copilot / Amp

  • Claude Code is described as more “fully agentic”: given a goal, it plans, edits, runs commands, tests, and can act like a junior dev, especially effective for small–medium, well-specified tasks.
  • Cursor is praised for deep IDE integration: inline completions, powerful TypeScript awareness, and Cursor Tab’s predictive editing; its agent mode is seen as less mature by some, but others report multi-minute autonomous runs with tool calls.
  • One view: Cursor heavily optimizes token/context to control cost, sometimes at quality’s expense; Claude Code is more liberal with context, often yielding stronger results but higher spend.
  • Several people find Claude Code more productive than Cursor (including in TS), others report the opposite, or that both are fine depending on language (Rails, etc.).
  • Comparisons with VSCode Copilot Agent (even when using Claude as backend) suggest Claude Code feels more transparent, controllable, and CLI-friendly, but this is strongly “you must try it yourself.”

IDE Ecosystem: VSCode & JetBrains

  • VSCode is seen as the primary target for agentic plugins; some IntelliJ users are considering switching.
  • Others report the JetBrains Claude Code plugin (though marked beta) now matches VSCode’s integration closely.
  • Some note IntelliJ-based IDEs still lag VSCode for broader agent/MCP integration and performance, while others criticize VSCode’s bloat and praise JetBrains.

Workflows, Agents & Git Worktrees

  • Many use Claude Code as a terminal agent inside their existing IDE (vim/tmux, JetBrains, or VSCode) rather than adopting a new editor like Cursor.
  • A subthread argues IDE integration is the wrong form: better to manage multiple git worktrees, each with its own Claude Code session/agent, and review diffs asynchronously—essentially an “agent management” IDE.
  • Others counter that IDE plugins can handle virtual branches and that swarm-of-agents workflows create context-switching and review burdens; some prefer a single focused agent plus manual iteration.
  • Rate limiting emerges as a practical constraint: some hit 429s when running multiple Claude sessions even on expensive plans; others run several in parallel without issue. Cost is framed as trivial for companies but significant for individuals or students.

Languages, Code Quality & LLM-Friendly Design

  • Users report Claude Code generates notably solid Go code with few hallucinations, while JS/TS remains more error-prone due to ecosystem complexity and clever idioms.
  • LSP integration reduces but doesn’t eliminate TypeScript errors; several wish Claude Code would always auto-run TS checks.
  • There is speculation that simpler, more explicit languages (Go, or even more constrained designs) are better for LLMs; some already choose languages (Go over Python, alternatives to Rust) partly because tools/LLMs handle them better.
  • Multiple comments note a shift toward designing codebases and documentation for agents: flatter structures, consistent patterns, more declarative metadata, and “write it like for a junior dev” instructions.

UX Details: Prompts, Diffing, Debugging

  • Users appreciate diff viewing directly in VSCode, no longer needing a side-by-side terminal + editor setup.
  • Some miss Cursor’s management of long prompts; Claude Code’s terminal UI can discard a long draft on an errant keypress, though history navigation, saved sessions, and using external files as “plans” are workarounds.
  • Desired future features include stronger debugging integration (stack inspection, variable viewing) and better notification/coordination UX for agents attached to multiple branches or worktrees.

Privacy, Trust & Policy Perception

  • One commenter argues that using Claude Code on proprietary code is inherently risky and “competes” with Anthropic’s goals, implying privacy is illusory.
  • Others respond that paid Anthropic products (Pro/API/Claude Code) do not use customer data for training and keep transcripts only briefly (e.g., 30 days) with explicit safeguards; they see strong corporate adoption as evidence that data is not being harvested.

Competition & Pricing Sentiment

  • Some are glad to see Claude Code as a competitive alternative after GitHub Copilot introduced premium request limits without price cuts; this triggered cancellations and frustration around “metered thinking.”
  • Claude Code is acknowledged as expensive—one task can burn through noticeable credit—but high-end subscriptions are seen as effectively “unlimited” in practice. There’s concern that current pricing may be subsidized and not sustainable long-term.