Zed (with AI)

What it is and how it works

Zed is a native code editor built in Rust by the team that originally created Atom (and, before that, worked on Eclipse-era tooling at TextMate's lineage). Where Atom and VS Code are Electron applications — a Chromium renderer plus Node.js wrapped around a text buffer — Zed is a from-scratch GPU-accelerated application. It renders its UI directly through a custom GPU rendering pipeline (originally built on Metal for macOS, with GPUI as the underlying UI framework the Zed team also open-sourced) rather than delegating layout and paint to a browser engine. That architectural choice is the whole reason Zed exists: the founders wanted an editor that felt as responsive as a native application, with sub-frame input latency, even on very large files and with many panes open.

The AI layer sits on top of that native core rather than replacing it. Zed ships an AI panel (a persistent chat/agent surface) and an inline assistant (transform-selected-text-in-place, à la Copilot's inline edit or Cursor's Cmd-K). Both are provider-agnostic: Zed does not bundle a proprietary model or require a Zed-hosted inference backend. You configure API keys for Anthropic, OpenAI, Google, or a self-hosted/local endpoint (Ollama, LM Studio, or any OpenAI-compatible server) and Zed routes requests directly from the client to that provider. This is a meaningfully different trust model from AI-native forks like Cursor or Windsurf, which typically proxy model calls through their own backend for context assembly, telemetry, and (in some cases) model routing/caching. Zed's BYO-key design means your code and prompts go straight from your machine to the provider you chose, with no intermediate hop owned by Zed Industries — a detail that matters to teams with strict data-handling policies.

Zed also supports the Language Server Protocol and, more recently, the Model Context Protocol (MCP) as an extension point, letting the AI panel call out to external tools and context providers rather than only working off open-buffer context. Collaboration (multiplayer editing, shared terminals) is a separate, longstanding Zed feature that predates the AI push and remains one of its differentiators.

Deployment patterns

For a solo developer, Zed is a single binary download (macOS and Linux have been the primary platforms since launch; Windows support arrived later and is comparatively newer, so expect a few more rough edges there) with essentially zero setup: install, sign in optionally for collaboration/theme sync, drop in an API key for whichever model provider you use, and start editing. There's no project-level daemon or background indexing service to manage the way there is with, say, a heavyweight JetBrains IDE.

The "BYO API key" model shapes the homelab/local-inference pattern specifically: because Zed doesn't force you through its own hosted model, pointing it at a local llama.cpp server, Ollama instance, or vLLM/TGI endpoint that exposes an OpenAI-compatible /v1/chat/completions route is a supported, first-class path rather than a hack. This is the pattern most relevant to a local-AI-focused audience — you get Zed's native editing performance with inference happening entirely on hardware you control, no cloud round-trip at all if you don't want one.

For teams, Zed's collaboration features (shared editing sessions, voice/text chat in the editor) plus centrally-issued API keys (via a shared vault or per-seat provider accounts) are the typical setup; there is no Zed-run enterprise control plane comparable to what GitHub Copilot for Business or Cursor's team admin console offers, so key management and usage governance are left to whatever the team already uses (1Password, environment injection via CI/CD secrets, etc.).

How it compares

Against VS Code plus Copilot or Continue.dev, Zed's pitch is raw editing performance and a smaller, more auditable codebase (VS Code plus its extension host is a much larger surface). The tradeoff is real: VS Code's plugin ecosystem is an order of magnitude larger, and most language-specific tooling, debuggers, and niche integrations land there first or exclusively. Zed's extension system is newer and narrower.

Against Cursor, which is a VS Code fork with deep AI-native features (multi-file agent edits, codebase-wide semantic indexing, its own model routing/caching layer), Zed is more minimal and more transparent about where your data goes, but Cursor's agent mode is generally more capable out of the box for large, multi-file autonomous edits — Cursor was built around that workflow from day one, while Zed's AI features feel more like a well-integrated add-on to an already-excellent editor.

Against JetBrains IDEs with AI Assistant, Zed is dramatically lighter weight and faster to open, but gives up JetBrains' deep language-specific static analysis, refactoring tools, and mature debugger integrations that took over a decade to build for Java, Kotlin, and friends.

Best use cases and honest limitations

Zed is a strong fit for developers who prioritize raw input latency and native feel, who want multi-provider flexibility (including pointing at local model servers) without a vendor-controlled AI backend in the loop, and who work primarily in mainstream languages where Zed's LSP support is solid (Rust, Python, TypeScript/JavaScript, Go). Its open-source nature also makes it a reasonable choice for anyone auditing exactly what an editor does with their code before AI requests leave the machine.

It's a weaker fit if you depend on a specific VS Code extension with no Zed equivalent, if you need JetBrains-grade refactoring/debugging for a large enterprise Java or C# codebase, or if you want an agent that autonomously drives large multi-file changes with minimal hand-holding — that's still more mature in Cursor and similar AI-native forks. Windows users should also weigh that support there is newer than on macOS/Linux. Pricing-wise, Zed itself is free and open source; the only cost is whatever API usage you rack up with your chosen model provider, which is a plus for cost transparency but means there's no bundled "unlimited AI" tier the way some subscription-based competitors offer.