overview.md

AI

Zed integrates AI throughout the editor: agentic coding, inline transformations, edit prediction, and direct model conversations.

How Zed approaches AI

Zed's AI features run inside a native, GPU-accelerated application built in Rust. There is no Electron layer between you and the model output.

  • Open source. The editor and all AI features are open source. You can read how AI is implemented, how data flows to providers, and how tool calls execute.
  • Multi-model. Use Zed's hosted models or bring your own API keys from Anthropic, OpenAI, Google, Ollama, and 8+ other providers. Run local models, connect to cloud APIs, or mix both. Switch models per task.
  • External agents. Run Claude Agent, Gemini CLI, Codex, and other CLI-based agents directly in Zed through the Agent Client Protocol. See External Agents.
  • Privacy by default. AI data sharing is opt-in. When you use your own API keys, Zed maintains zero-data retention agreements with providers. See Privacy and Security.

Agentic editing

The Agent Panel is where you work with AI agents. Agents can read files, edit code, run terminal commands, search the web, and access diagnostics through built-in tools.

You can extend agents with additional tools through MCP servers, control what they can access with tool permissions, and shape their behavior with rules.

The Inline Assistant works differently: select code or a terminal command, describe what you want, and the model rewrites the selection in place. It works with multiple cursors.

Code completions

Edit Prediction provides AI code completions on every keystroke. Each keypress sends a request to the prediction provider, which returns single or multi-line suggestions you accept with tab.

The default provider is Zeta, Zed's open-source model trained on open data. You can also use GitHub Copilot, Supermaven, or Codestral.

Text threads

Text Threads are conversations with models inside any buffer. They work like a regular editor with your keybindings, multiple cursors, and standard editing features. Content is organized into message blocks with roles (You, Assistant, System).

Getting started

Agentic Editing

  • Agent Panel: Chat with AI agents that can read, write, and run code in your project.

  • Rules: Define specific instructions for AI behavior.

  • Tools: The built-in capabilities agents use: file operations, terminal commands, web search.

  • Tool Permissions: Configure granular permission rules for agent tool actions.

  • Model Context Protocol: Extend agents with custom tools via MCP servers.

  • Inline Assistant: Transform selected code or terminal output with ctrl-enter.

Edit Prediction

  • Edit Prediction: AI-powered autocomplete that predicts multi-line edits as you type.

Text Threads

  • Text Threads: Lightweight conversations with models inside any buffer.