Zed AI

What it does

• Agentic Editing: the agent discovers context, edits files, and iterates until the task is done — at the editor's native speed (Zed is written in Rust)
• Review-first workflow: agent work lands in an editable unified diff; accept or reject all changes before anything is committed
• Any agent via ACP: Claude Agent, Codex, OpenCode, Gemini CLI and more connect to the same agent panel; Zed, the Zeta model, and ACP are all open source
• MCP servers extend agent knowledge and tools; tool permissions are configurable
• Real-time collaboration: follow the agent live on the same multiplayer infrastructure Zed uses for humans

When to use it

For developers who want an agent inside the editor rather than a separate terminal app: hand off a task, watch it navigate the codebase in real time, then review the whole change as one diff before committing. Teams standardizing on one editor across Claude, Codex, or OpenCode backends get all of them through one panel.

Getting started

Agent features ship inside the editor — no plugin to add:

1. Download Zed from https://zed.dev/download (macOS/Linux)
2. Open the Agent Panel and sign in to Zed AI, or bring your own keys (Anthropic, OpenAI, Google and others) per https://zed.dev/docs/ai/llm-providers
3. To drive external agents, install the agent's ACP adapter — see https://zed.dev/acp for the directory

Things to know

Capabilities depend on which agent/provider you connect: model access, rate limits, and tool behavior follow the backend, not Zed. Zed also ships its own Zeta model alongside the ACP agents — the agent metrics page on zed.dev/ai shows real usage split across Claude Agent, Zed Agent, Codex CLI and others.