[DRAFT] Zed: The AI editor everyone is talking about

Everyone on my timeline is switching to Zed. Before I install it and blow up my Claude Code workflow, I wanted to know what's actually different — not the marketing, the real tradeoffs.

Here's what the research turned up.

What Zed actually is

Zed is a code editor written in Rust with a custom GPU-accelerated UI framework called GPUI. It renders directly to Metal on macOS (Vulkan elsewhere), the same way a video game renders — no web engine, no Electron, no abstraction layers between the text cursor and the GPU. That's the technical reason it's fast, and the benchmarks back it up: 0.12s startup vs VS Code's 1.2s, 222MB memory vs 3,549MB, ~2ms input latency vs ~25ms.

That last number is the one that matters. 25ms of input latency is below the perceptual threshold for most people, but above it for some. If you've ever felt VS Code or Cursor getting "heavy" under load, Zed is genuinely solving a different problem.

The AI features, concretely

There are two modes:

Inline Assistant — select code, hit Ctrl+Enter, describe the change, get an in-place rewrite. Works with multiple cursors simultaneously. This is the quick-edit path.

Agent Panel — conversational AI that can read files, write across the codebase, run shell commands, and call MCP tools. Three sub-modes: Ask (read-only), Write (edits), and Thread (persistent sessions that survive window restarts). Tool calls surface inline so you can inspect them.

The headline feature from Zed 1.0 (April 2026) is Parallel Agents — multiple agent threads running concurrently in the same window, each on different files or repos, each optionally using different models. You open the Threads Sidebar and run them side by side. This is architecturally novel. I haven't seen this done cleanly in any other editor.

Model support is broad: Claude, GPT, Gemini, DeepSeek, Bedrock, Ollama, LM Studio, OpenRouter. You can bring your own keys or use Zed's hosted credit ($5/month included with Pro). The "not paying Zed" path is fully supported.

Where Claude Code fits in

This is the most interesting part for me. Zed integrates Claude Code via the Agent Client Protocol — Claude Code runs as an external process, communicates over ACP via stdin/stdout, and appears inside Zed's editor with full syntax highlighting and diff review. You get the visual layer on top of the CLI tool.

The catch: there's a documented context window issue. When running Claude Code through Zed's ACP adapter, context gets compacted aggressively — reportedly to 120–200K tokens instead of Claude's actual 1M maximum. After compaction the model forgets what it read and changed earlier in the session. That's a real limitation, tracked as an open GitHub issue with no fix yet.

The other missing pieces: not all slash commands work inside Zed's agent window, agent teams aren't supported, hooks aren't supported, and there's no checkpointing.

So the Zed + Claude Code combo gives you visual diffs and syntax-highlighted edits inside a fast editor, but you trade away context depth and some command functionality compared to running Claude Code in the terminal.

Zed vs Cursor vs Claude Code terminal

The honest comparison I kept coming back to:

Cursor is still best-in-class for integrated AI editing. Composer does coordinated multi-file diffs better than anything Zed offers natively. If AI-first IDE experience is the goal, Cursor wins today.

Claude Code in the terminal is the right tool for complex autonomous tasks — full 1M context, full command support, no ACP wrapper overhead. The terminal isn't a limitation, it's the unobstructed path.

Zed is the right choice if the editor itself has become the bottleneck — if you're losing time to startup, memory pressure, or input lag. You still get functional AI (inline assistant, agent panel, Claude Code via ACP), just with the tradeoffs above.

The hybrid stack most engineers seem to land on: Zed for the editor, Claude Code in the terminal for heavy lifting. Or Cursor daily + Claude Code for the long autonomous runs.

What people actually say

Praise: the speed claims are real, parallel agents are genuinely useful once you understand the model, and the 1.0 release cleared a reliability bar that earlier versions didn't meet.

Complaints worth noting: CPU spikes in some workflows (not universal, but reported enough to be real), language server instability that's improving but not gone, and agent UX that's less mature than Cursor's — the distinction between agent threads and text threads confuses people, and context window behavior isn't always predictable.

The "LLM infested" complaint exists too. Some Zed users came for the fast editor and don't want AI at all. That's valid, and Zed supports disabling AI entirely.

What I haven't done yet

I haven't actually installed it. That's the next step — try the Inline Assistant on a real codebase, run the Agent Panel on something non-trivial, test the Claude Code ACP integration and see how the context compaction actually feels in practice.

The research says: fast, honest about limitations, parallel agents are the real differentiator, Claude Code integration is functional but degraded compared to terminal. That's a reasonable starting hypothesis. I'll update this post after I've run it for a few days.

Research done before first install — will update with hands-on experience.