Claude Code started as a chat interface with file access. You’d type, it’d respond, maybe edit some code. Simple.

That’s not what it is anymore.

The changes are quiet. No single release transforms the tool. But scroll through the release notes and you see a pattern: Claude Code is expanding in two directions simultaneously. First, it’s reaching into new domains: browser automation, language server integration, terminal configuration. Second, it’s polishing the small stuff: prompt suggestions, syntax highlighting, keyboard shortcuts.

Neither direction matters much on its own. Together, they compound.

Expanding Scope

On-Demand Tool Loading

The biggest change isn’t a new feature - it’s infrastructure. Claude Code now loads MCP tool definitions on-demand instead of upfront.

I wrote about Tool Search when Anthropic shipped it for the API in November. Back then, Claude Code still loaded every MCP schema at startup. A five-server setup could consume 50k+ tokens before you typed anything.

Finally shipped

As of this week, on-demand loading is live in Claude Code. Run /context and you’ll see MCP tools listed as “Available” without token counts. Not yet in the release notes, but it’s there.

The GitHub issue requesting this showed the problem: MCP servers consuming 54% of the 200k context at startup. With on-demand loading, that overhead drops to near zero until you actually invoke the tools.

This changes everything else in this post. Chrome automation, LSP, any MCP-based capability - you can have them configured without paying the context tax until you use them.

Chrome Automation

The Claude in Chrome extension lets you control a browser from the terminal. Navigate to a URL, click elements, fill forms, take screenshots. All without leaving your flow.

For testing UI, debugging frontend issues, or automating repetitive form tasks, it’s genuinely useful. You describe what you want to happen, Claude does the clicking.

The Chrome MCP schema is 14.1k tokens - but with on-demand loading, that cost only hits when you invoke browser automation. Available when needed, invisible when not.

LSP Integration

Claude Code now has access to language server features: go-to-definition, find references, hover documentation. The same code intelligence your IDE uses.

What this means in practice: Claude can follow the code semantically instead of grep-ing through text. When it needs to understand how a function is used, it can find all call sites instantly. When it’s confused about a type, it can look up the definition. The performance argument is compelling: 45 seconds of grep searches vs 50ms with LSP.

The cost? ~500 tokens added to system tools. Worth it for navigating unfamiliar codebases.

Early stage

LSP support is new and still rough. Enable with ENABLE_LSP_TOOL=1 environment variable. Expect bugs across different language servers.

Ergonomic Polish

Prompt Suggestions

As you type, Claude Code suggests completions. Start typing / and it surfaces relevant commands. Start a common prompt pattern and it offers to complete it.

Small feature. But fewer keystrokes per session adds up. Over weeks of use, the time saved compounds.

Syntax Highlighting and Terminal Setup

The syntax highlighting engine got a rewrite. Code in responses now renders with proper coloring. Toggle it with ctrl+t in the theme picker.

More useful: /terminal-setup now supports Warp, Kitty, Alacritty, and Zed. One command configures your terminal for optimal Claude Code experience. Shift+Enter for multi-line input, proper keybindings, consistent behavior.

The best tools adapt to your workflow. You shouldn’t have to adapt to them.

Before this, you’d manually configure each terminal. Now it just works.

The Remaining Overhead

On-demand loading fixes MCP bloat, but context costs don’t disappear entirely. Run /context and look at the breakdown:

  • System tools: ~17k tokens (includes LSP)
  • Autocompact buffer: 45k tokens reserved for summarization
  • Memory files: 4-5k tokens depending on your CLAUDE.md
  • Custom agents and skills: Variable

With everything enabled, you’re still working with ~130k of your 200k before typing anything. The difference: that overhead is now fixed costs for core functionality, not variable costs for tools you might not use.

What This Means

Claude Code is becoming a complete development environment. Browser automation, code intelligence, terminal integration, ergonomic polish. And now, on-demand tool loading so you don’t pay the context tax until you need the capability.

The details compound. The trade-offs finally compound less.

Coming soon

One more improvement in the pipeline: a complete rewrite of the terminal renderer that reduces flickering by 85%. Rolled back for holiday stability, but when it returns, the strobe-light effect during long sessions should finally be gone.

Related Posts