Zed: The Open Source Agentic IDE
What Makes Zed Different
Zed is not another Electron-based editor. It's built from the ground up in Rust, which means real performance without the memory bloat that plagues other IDEs. If you've ever hit a "window unresponsive" error while running multiple projects, you understand why this matters.
The bigger story is the Agent Client Protocol—an open standard that decouples your editor from any single AI provider.
The Agent Client Protocol Explained
The protocol standardizes communication between code editors and AI agents. Without it, every new agent-editor combination requires custom integration work. You're locked into whatever the editor's creators decided to support.
Zed's approach flips this. You can run Claude Code, Codex, or Gemini CLI through the same interface, using your existing subscriptions. When a new model drops—say, Gemini 3—you don't wait for an update. You switch agents in a new thread and keep working.
This standard is gaining traction beyond Zed. Augment Code's Auggie and JetBrains have adopted it. Open source tooling that benefits competitors is rare. It happens when the creators prioritize user flexibility over ecosystem lock-in.
Getting Started
Installation is straightforward. Zed runs on macOS, Linux, and Windows. The repository is open source—star it if you use it.
Key bindings will feel familiar if you're coming from VS Code or Cursor. Open sidebars and terminals with the same shortcuts. The agent panel sits on the right, ready when you need it.
Agent Integration in Practice
Starting a conversation with an agent works like running a CLI command, but inside the IDE. Select your agent—Claude Code, Codex, whatever—and Zed spins it up in a new thread. You get the same performance as the terminal version, but with a structured UI that tracks changes visually.
The interface shows exactly what the agent is doing: which files it's reading, what commands it's running, and how it understands your project structure. No token streaming clutter. No performative "look how fast I am" animations. Just a clean list of actions you can follow or review later.
Context and Control
You have multiple ways to steer the agent:
- @ mentions for specific files, symbols, or previous conversation threads
- Rules for consistent behavior across sessions
- Web fetch for external documentation or research
- MCP servers for extended capabilities like Firecrawl search
Permission levels let you control how autonomous the agent behaves. "Ask" mode requires confirmation for every action. "Bypass" mode lets the agent run freely—useful for low-stakes refactors or when you trust the context and instructions.
Building with Agents: A Real Example
The demo walks through building a Next.js application. The user requests a neo-brutalist homepage with black and white as primary colors. Claude Code generates the implementation, but the interaction reveals something more interesting.
When asked to research and write blog posts about GPT 5.1, Gemini 3, and Sonnet 4.5, the agent pauses. It found solid information on GPT 5.1, but flagged that Gemini 3.5 lacks credible sources. Rather than hallucinate content, it asks for clarification. This kind of transparency—admitting knowledge limits instead of generating plausible-sounding falsehoods—is exactly what you want from an AI assistant.
The resulting blog post includes properly formatted tables, source citations, and a cohesive design that matches the neo-brutalist aesthetic. All generated through iterative file edits you can track in real-time.
Why This Matters
The CLI-first trend in AI coding tools has merit. Terminal environments are fast and familiar. But professional development often benefits from IDE features: integrated debugging, file trees, and visual diff views. Zed gives you both—the raw capability of agentic CLI tools within a structured, performant editing environment.
You keep your workflow when switching between Claude Code and Codex. The keyboard shortcuts stay the same. The project context persists. Only the underlying model changes.
As model capabilities continue leapfrogging each other—one week it's GPT, the next it's Claude, then Gemini—this flexibility becomes essential. You're not rebuilding your development environment every time you want to try a new agent. You're just opening a new thread.
