# Zed A high-performance, open source code editor with a built-in Agent Panel, MCP support, and ACP-based external agents. Schema: 1.0 · Data: ee5de863c30bdbe4fe7dd3e07c7e22ea9bf4ab2248dbc7c35cdbf60bc28be1db ## Configuration ### MCP Status: supported Configure MCP servers (context servers) in settings.json under the context_servers key, or install them as Zed extensions. Supports local stdio and remote servers with OAuth. - global: `~/.config/zed/settings.json` (json) — key: context_servers — macOS/Linux. On Windows: %APPDATA%/Zed/settings.json. - project: `.zed/settings.json` (json) — key: context_servers - `source`: "custom" | "extension" — How the server is provided. Entries added via the UI set 'custom'; extension-provided servers use 'extension'. - `command`: string — Command to start a local (stdio) server. - `args`: string[] — Arguments passed to the command. - `env`: record — Environment variables for a local server. - `url`: string — Endpoint URL for a remote server. - `headers`: record — HTTP headers for a remote server (e.g. Authorization: Bearer ). - transports: stdio, http - auth: oauth, headers - Tools: supported - Prompts: supported - Resources: unsupported — Official docs state Zed currently supports MCP's Tools and Prompts features. - Roots: unsupported — Not listed among supported MCP features in official docs. - Elicitation: unsupported — Docs welcome contributions for Elicitation support. - Sampling: unsupported — Docs welcome contributions for Sampling support. - MCP servers can also be installed as Zed extensions from the extension registry (Settings > AI > MCP Servers > Install from Extensions). - Remote servers without an Authorization header trigger the standard MCP OAuth flow. - Zed handles notifications/tools/list_changed and reloads a server's tool list at runtime without a restart. - Servers configured in Zed are forwarded to External Agents over the Agent Client Protocol (ACP). - Per-tool approval uses agent.tool_permissions with the key format mcp::; agent.tool_permissions.default is one of confirm, allow, or deny. - Custom agent profiles can scope which context servers and tools are enabled via enable_all_context_servers and per-server tool toggles. Local and remote context servers ```json { "context_servers": { "local-mcp-server": { "command": "some-command", "args": ["arg-1", "arg-2"], "env": {} }, "remote-mcp-server": { "url": "https://example.com/mcp", "headers": { "Authorization": "Bearer " } }, "remote-mcp-server-with-oauth": { "url": "https://mcp.example.com/mcp" } } } ``` Source: https://zed.dev/docs/ai/mcp Checked: 2026-07-12 ### Skills Status: supported SKILL.md skill folders following the open Agent Skills standard, loaded autonomously from a catalog or invoked manually via slash command or @-mention. - project: `.agents/skills//SKILL.md` (markdown) — Project-local skills, only loaded from trusted worktrees. - global: `~/.agents/skills//SKILL.md` (markdown) — Global skills, available in every project. - `name`: string (required) — Lowercase letters, numbers, and hyphens only; 1-64 chars; no leading/trailing/consecutive hyphens. Should match the folder name. - `description`: string (required) — What the skill does and when to use it. Keep under 1024 bytes; longer descriptions load with a warning. - `disable-model-invocation`: boolean — When true, hides the skill from the agent's catalog; it remains invocable via slash command or @-mention. - invocation: /skill-name, @skill, automatic - Skills replaced Zed's reusable Rules Library. - Flat layout only: skills must be direct children of the skills root; nested folders are not discovered. - Project-local skills only load from trusted worktrees; a project-local skill overrides a global skill with the same name. - Live reload: SKILL.md edits take effect immediately without restarting the session. - The total size of all skill names and descriptions is capped at 50KB for the catalog. - Skills can bundle scripts/, references/, and assets/ that the agent loads on demand. - Skills can be shared via self-contained zed://skill links and imported from GitHub .md URLs; a built-in /create-skill skill scaffolds new skills. - Invoking a user-created skill prompts for permission (same flow as tool permissions); built-in skills do not prompt. Manual-only deploy skill ```markdown --- name: deploy description: Deploy the current branch to production. disable-model-invocation: true --- ## Instructions Step-by-step instructions for the agent... ``` Source: https://zed.dev/docs/ai/skills Checked: 2026-07-12 ### Rules Status: supported Always-on instructions via AGENTS.md: a personal global file plus one project instruction file, with compatibility for .rules, .cursorrules, CLAUDE.md, and similar. - global: `~/.config/zed/AGENTS.md` (markdown) — Personal instructions for every project. On Windows: %APPDATA%/Zed/AGENTS.md. - project: `AGENTS.md` (markdown) — Project instructions at the worktree root. First matching compatibility file wins (see notes). - project: `.rules` (text) — Zed-native compatibility instruction file; highest priority in the match order. - Zed's Rules Library was replaced: reusable rules became Skills, always-on rules became Instructions (AGENTS.md). - For project instructions Zed uses the first matching file in this order: .rules, .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md, AGENT.md, AGENTS.md, CLAUDE.md, GEMINI.md. - Project instructions override personal AGENTS.md when they conflict. - Instructions apply to the Zed Agent; External Agents and Terminal Threads read their own native instruction files. Project AGENTS.md ```markdown # Project instructions - Run `cargo test` before committing. - Follow the repository's error-handling conventions in src/errors.rs. ``` Source: https://zed.dev/docs/ai/instructions Checked: 2026-07-12 ### Hooks Status: unsupported Zed has no lifecycle hook system for the Agent Panel: no user-configurable scripts fire on agent events such as tool use, prompt submit, or session start/end. - Agent lifecycle hooks (custom commands, pre/post tool-use scripts) are open feature requests on the Zed GitHub repo (e.g. issues #57890, #52688) but are not shipped. - The unrelated create_worktree Task hook runs when Zed creates a linked git worktree for parallel agents; it is a workspace task hook, not an agent lifecycle hook. Checked: 2026-07-12 ### Commands Status: partial Slash commands in the Agent Panel invoke skills by name, plus the built-in /compact. There is no separate custom prompt-template command format; custom commands are authored as skills. - project: `.agents/skills//SKILL.md` (markdown) — Skills double as slash commands; see the skills surface for the full format. - global: `~/.agents/skills//SKILL.md` (markdown) — Global skills are also available as slash commands in every project. - invocation: /skill-name, /compact, /create-skill - Typing / in the Agent Panel message editor lists installed skills; selecting one injects the skill's instructions as context. - /compact manually compacts the current Zed Agent thread. - /create-skill invokes the built-in skill-creation skill. - Legacy Text Threads have their own built-in context slash commands (/file, /fetch, /prompt, /now, /diagnostics, /symbols, /tab, /selection, /terminal, /default); these do not work in the modern Agent Panel. See https://zed.dev/docs/ai/text-threads. - User-defined prompt-template command files (e.g. a .zed/commands/ directory) are an open feature request, not a shipped feature. Source: https://zed.dev/docs/ai/skills Checked: 2026-07-12 ### Settings Status: supported Agent behavior is configured in settings.json under the agent key: tool permissions, profiles, auto-compaction, feature-specific models, and model parameters. - global: `~/.config/zed/settings.json` (json) — key: agent — macOS/Linux. On Windows: %APPDATA%/Zed/settings.json. - project: `.zed/settings.json` (json) — key: agent — Project settings override user settings for this worktree. - `tool_permissions.default`: "confirm" | "allow" | "deny" — Default tool-approval behavior for the Zed Agent. confirm (default) prompts before every tool action. - `profiles`: object — Named agent profiles that toggle built-in tools and context servers (e.g. enable_all_context_servers, per-server tools). - `auto_compact`: { enabled: boolean, threshold: string | number } — Automatic thread compaction. Threshold is a percentage string ("90%"), a token count, or a negative remaining-token count. - `model_parameters`: array — Temperature overrides per provider, model, or provider/model pair; matching entries are checked last to first. - `inline_assistant_model`: object — Model override for the Inline Assistant. - `commit_message_model`: object — Model override for Git commit message generation. - `thread_summary_model`: object — Model override for thread summaries. - `subagent_model`: object — Model override for subagents. - `commit_message_instructions`: string — Instructions applied only to generated Git commit messages. - The Settings Editor (agent: open settings) exposes LLM providers, External Agents, and MCP servers; zed: open settings file opens the raw JSON. - MCP servers live in the same settings.json under the context_servers key (see the mcp surface). - In Zed v0.224.0+ tool approval is controlled by agent.tool_permissions.default; earlier versions used the agent.always_allow_tool_actions boolean (default false). - Per-tool permission rules support MCP tools with the key format mcp::. Auto-compaction and commit message instructions ```json { "agent": { "auto_compact": { "enabled": true, "threshold": "90%" }, "commit_message_instructions": "Use the Conventional Commits format: (): ." } } ``` Source: https://zed.dev/docs/ai/agent-settings Checked: 2026-07-12 ## Search, browser & identification Not researched. Unknown does not mean unsupported.