Three shapes, one protocol, and Claude Code's 30GB memory bill

AI coding tools have settled into 3 categories, all speaking the same protocol. Pick the shape that fits your day — and check your RAM before running parallel agents.

The AI coding tool market has settled into 3 distinct shapes, and as of May 2026, choosing between them is less about capability and more about how you want to work. Anthropic just shipped a redesigned Claude Code desktop app built for parallel tasks, while developer Simon Willison flagged that those same Claude processes consumed ~30GB of memory on his Mac. The tools are getting more capable and more expensive — in compute, not dollars.

A useful framing comes from a recent guide on agensi.io: terminal-based agents like Claude Code, Codex CLI, and Gemini CLI; full IDEs like Cursor and Windsurf; and editor extensions like GitHub Copilot. This isn't marketing taxonomy — it maps to how you actually work. Terminal agents suit developers who live in shells and want autonomous multi-step execution. IDEs suit those who want AI woven into a full editor with file trees and diffs. Extensions suit teams already committed to VS Code or JetBrains who want suggestions without changing their environment.

Underneath all three, the Model Context Protocol has gone universal. As of May 2026, Anthropic, OpenAI, Google, and Microsoft have integrated native MCP support into their flagship models, and frameworks like LangChain and LlamaIndex have moved MCP from experimental to default for tool-calling. The practical effect: a custom MCP server you write today plugs into Cursor, Claude Code, and Continue without per-IDE adapters. Vendor lock-in at the tool-integration layer is melting fast.

The cost is no longer dollars but resources. Anthropic's May 2026 Claude Code redesign adds a session sidebar, drag-and-drop workspace, integrated terminal and file editor, and SSH support on Mac — explicitly built for running more tasks in parallel. But Willison noted on X that idle Claude Code processes across his terminal windows ate roughly 30GB of memory, with one session alone at 4.9GB. If you run a 16GB MacBook, parallel agentic coding is now a workstation decision, not a casual one.

For builders shipping production code, this means picking a category deliberately and budgeting machine resources accordingly. For buyers, MCP universality cuts the cost of switching tools next quarter. For teams choosing today, the question isn't which tool wins — it's which shape fits the day you actually have.

The bottom line: Pick the AI coding tool shape that matches how you already work, and check your machine's RAM before committing to parallel agents.

Sources