Zed AI: Zed Agent, external agents, skills, edit prediction and pricing

What is Zed AI?

Zed AI is the set of AI features built into Zed, the open source code editor: a native coding agent, support for outside agents like Claude Code and Codex, an inline assistant, and AI code completions called edit predictions. You can use it with Zed-hosted models, your own API keys, some existing subscriptions or local models.

Zed's GitHub repository describes the editor as "a high-performance, multiplayer code editor from the creators of Atom and Tree-sitter." The AI overview groups the AI features into three areas:

• Agents: how agentic work runs, through the Zed Agent, External Agents or Terminal Threads.
• Model access: how Zed reaches models, through Zed-hosted models, API keys, subscriptions, gateways or local models.
• Features: the Agent Panel, Parallel Agents, the Inline Assistant, Edit Prediction and Git commit message generation.

This guide follows the Zed docs as read on 26 October 2026. The latest stable release on GitHub was v1.25.3, published 23 October 2026.

What are the three ways to run an agent in Zed?

Zed gives you three agent paths, and knowing which one you are using explains most of its settings:

The Zed Agent page says the native agent can read and search your project, edit files, run terminal commands, use MCP tools, follow Agent Profiles and show changes in Zed's review UI. Settings for the Zed Agent do not carry over to the other two paths. The Terminal Threads docs say that Zed Agent profiles, tool permissions, skills and MCP settings "do not automatically apply to Terminal Threads."

How do you use the Agent Panel?

The Agent Panel is where you prompt agents. Open it with agent: new thread from the Command Palette or the sparkle icon in the status bar, then configure a model for the Zed Agent or install an external agent.

Features worth knowing:

• Context with @. Type @ to mention files, directories, symbols, previous threads, skills, diagnostics, branch diffs or URLs to fetch. You can also drag images in.
• Checkpoints. After the agent edits code, a Restore Checkpoint button returns the project to its state before that message.
• Reviewing changes. Click Review Changes to open every edit in one multibuffer tab, then accept or reject each hunk or the whole set.
• Queueing and steering. Messages you send while the agent works are queued, and with the Zed Agent you can toggle Steer to have one interrupt at the next step.
• New From Summary. Start a fresh thread seeded with a summary of the current one when you near the context limit.

Agent Profiles decide which tools the Zed Agent can use. The profiles page lists three built-in profiles: Write for reading, editing and running commands, Ask for read-only questions, and Minimal with no project tools. Whether a tool call runs without asking is a separate setting, agent.tool_permissions, where regex patterns can always allow commands such as cargo test, always confirm risky ones such as sudo, or block them.

How do you use Claude Code, Codex and other agents in Zed?

Install them as External Agents. The external agents page says the ACP Registry "is the primary way to install common External Agents in Zed." Open it with zed: acp registry, or go to Agent Settings, click Add Agent and choose Install from Registry. The agent then appears in the new-thread menu.

Each external agent handles its own sign-in. For example, Claude Agent uses /login to pick between an API key and Claude Code authentication, and an Anthropic key you configured for the Zed Agent does not carry over. The docs also note that "Zed does not charge for External Agents." Billing is between you and the agent's provider.

If an agent has no ACP integration, open a Terminal Thread from the new-thread menu and run its CLI there. Setting agent.terminal_init_command to "claude", for example, starts that command in every new Terminal Thread.

With Parallel Agents you can run several threads at once from the Threads Sidebar, each with its own agent, and isolate one in a separate Git worktree when two threads might touch the same files.

How do you give the Zed Agent project rules?

Through Instructions, which are always-on context. The instructions page says "Zed supports AGENTS.md as the primary instruction file." Personal instructions go in ~/.config/zed/AGENTS.md. For a project, Zed uses the first file it finds in this order:

1. .rules
2. .cursorrules
3. .windsurfrules
4. .clinerules
5. .github/copilot-instructions.md
6. AGENT.md
7. AGENTS.md
8. CLAUDE.md
9. GEMINI.md

Project instructions override personal ones when they conflict. Because only the first match is used, a repo with both .cursorrules and AGENTS.md gives the Zed Agent the .cursorrules file. Zed's older Rules feature has been replaced: on-demand rules became Skills and always-on rules became personal AGENTS.md.

Does Zed support Agent Skills?

Yes. The skills page defines a skill as "a folder containing a SKILL.md file with metadata and instructions." The agent sees a catalog of every installed skill's name and description and loads one when a task matches. You can also type / and pick a skill, or use @skill.

Skills must sit directly under the skills root, so nested group folders are not discovered. The name field must be lowercase letters, numbers and hyphens, max 64 characters, and description should stay within 1,024 characters.

Zed adds a few conveniences. The built-in /create-skill skill walks you through writing one, agent: create skill from url imports a SKILL.md from a GitHub link, and a zed://skill share link embeds a whole skill for a teammate to review and save. Adding disable-model-invocation: true to the frontmatter keeps the agent from triggering a skill on its own, which suits deploy or release workflows. When the agent invokes a skill you created or installed, Zed asks you to allow or deny it, using the same permission flow as other tools. Skills built into Zed do not prompt.

Since Zed uses .agents/skills/, the same folders work with other agents that follow the Agent Skills standard.

How do you add MCP servers to Zed?

Zed connects to MCP servers as "context servers." The MCP page says Zed supports MCP's Tools and Prompts features. You can install servers packaged as Zed extensions, such as Context7 or GitHub, or add your own under Settings > AI > MCP Servers with Add Local Server or Add Remote Server. Both write to context_servers in your settings:

{
  "context_servers": {
    "local-mcp-server": {
      "command": "some-command",
      "args": ["arg-1", "arg-2"],
      "env": {}
    },
    "remote-mcp-server": {
      "url": "https://example.com/mcp"
    }
  }
}

The Zed Agent uses these directly, and Zed can forward them to external agents over ACP.

What are the Inline Assistant and edit prediction?

The Inline Assistant rewrites a selection in place. Press ctrl-enter in an editor or the terminal panel, describe the change, and Zed replaces the selection with the model's answer. With multiple cursors the same prompt runs at every cursor, and you can configure it to send one prompt to several models at once. It uses Zed-configured models, not external agents.

Edit prediction is Zed's AI autocomplete: suggestions appear as you type and you accept them with tab. The default provider is Zeta, "an open source model developed by Zed," and you can switch to GitHub Copilot, Mercury Coder or Codestral.

How much does Zed AI cost?

As of October 2026, the Zed pricing page lists three plans:

Hosted usage on Pro is billed at "API list price +10%," and you can set a spending cap. There is a 14-day Pro trial with no credit card.

You can also bring subscriptions you already pay for. The subscriptions page says ChatGPT Plus and Pro work through Zed's ChatGPT Subscription provider, GitHub Copilot works for chat models and edit prediction, and Claude Pro or Max has "no direct Zed LLM provider path," so you use Claude Agent or Claude Code instead.

Where do you keep the prompts and skills you use in Zed?

Zed reads AGENTS.md and .agents/skills/, the same files many other agents use, so keeping them in one versioned place pays off. Kipt keeps prompts and Agent Skills as Markdown in a private GitHub repo you own, with {{name}} blanks and Cmd+K search. If you want to turn a prompt into a SKILL.md first, the free SKILL.md generator checks the name and description limits in your browser.