← Coding

ZedZed

Zed Industries

编程Coding IDEIDE 本地Local 协作Collaboration partial lifecycle: active 核验Checked 2026-10-23

Zed 在一个编辑器里编排三种 agent 路径：Zed 原生 Agent，Claude / Codex，OpenCode / Copilot / Cursor。

Not "yet another AI coding tool" but an orchestration layer for three agent paths — use Zed's native agent in the same editor, or attach Claude / Codex / OpenCode / Copilot / Cursor.

适合与不适合When to use it, when not to

适合重视编辑器性能（Rust）、解析准确度（Tree-sitter）与多人协作的场景。

也适合想在一个编辑器里混用原生 Agent 与外部 CLI Agent 的用户。这是它最独特的价值。

不适合把「编辑器自带 AI」当唯一卖点去找的用户。 Zed 的定位是 agent 的编排层，不是「我自带一个模型」。

也不适合需要实测才知道大仓库表现的人 —— 索引策略官方未公开说明。

Scenarios where editor performance (Rust), syntactic parsing accuracy (Tree-sitter) and multi-user collaboration matter. Users who want to mix the native agent with external CLI agents in one editor — this is its most distinctive value.

Not a fit for people looking for an editor whose only selling point is built-in AI — Zed positions itself as an orchestration layer for agents, not as "I ship my own model". Also not a fit if you need hands-on measurements of large-repository behaviour — the indexing strategy is not publicly documented.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网直连实测 200。official site measured 200 direct.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 官方列出 5 条模型接入路径（docs/src/ai/llm-providers.md）：

Zed-Hosted Models —— 经 Zed 计费
API Access —— 用自己的 provider 额度，含 Anthropic 兼容与 OpenAI 兼容两种
Existing Subscription —— 复用已付费的 ChatGPT / Claude / Copilot 等
Gateway —— 经 OpenRouter / Bedrock / Vercel 等网关路由
Local Model —— 本地运行或自托管

路径选择由用户决定，不是绑定的单一供应商。 llm-providers.md 逐条列明这五条。
The official docs list 5 model-access paths (docs/src/ai/llm-providers.md):

Zed-Hosted Models — billed through Zed
API Access — your own provider quota, including Anthropic-compatible and OpenAI-compatible
Existing Subscription — reuse a paid ChatGPT / Claude / Copilot etc.
Gateway — routed via gateways like OpenRouter / Bedrock / Vercel
Local Model — running locally or self-hosted

Path choice is up to the user, not a bound single provider. llm-providers.md lists these five item by item.
运行位置Where it runs: 本地桌面应用，用 Rust 写。

官方自述来自 Atom 与 Tree-sitter 的创造者，主打性能与多人协作。

核验记录：文档含 dev-containers.md（开发容器）。另有 toolchains.md（工具链管理）一页。

关闭应用即停止。
A local desktop application, written in Rust.

The official self-description: from the creators of Atom and Tree-sitter, leading with performance and multiplayer collaboration.

Verification record: the docs include dev-containers.md (dev containers). There is also a toolchains.md (toolchain management) page.

Closing the app stops it.
本地文件Local files: 直接读写工作区，支持多根目录（multi-root）。

也可经 SSH / WSL 打开远端文件夹，即 remote development。

文件树扫描由 file_scan_exclusions 与 project_panel.hide_gitignore 控制。符号能力来自 Tree-sitter 解析与语言服务器（LSP）。编辑预测另有独立 provider。

Agent 内置检索是四种：grep（正则）、find_path（glob）、list_directory、read_file。

官方未见代码库向量化语义索引（已查 Tools 页）。

大仓库的限制官方有明说：一次打开 >100,000 文件的超大目录表现不佳。官方建议只开具体项目或子目录（remote development 页）。
Reads and writes the workspace directly, with multi-root support.

Can also open remote folders via SSH / WSL, i.e. remote development.

File-tree scanning is controlled by file_scan_exclusions and project_panel.hide_gitignore. Symbol capability comes from Tree-sitter parsing and language servers (LSP). Edit prediction has its own separate provider.

The agent's built-in retrieval has four kinds: grep (regex), find_path (glob), list_directory, read_file.

No codebase vectorized semantic index found officially (Tools page checked).

The official docs state the large-repo limit explicitly: huge directories with >100,000 files open at once perform poorly. The official advice is to open only specific projects or subdirectories (remote development page).
关机后的任务Tasks after shutdown: 本地桌面应用。agent 线程在本地进程内跑。远端主机可以是自建的 SSH / WSL 主机。

绑定客户端在线。 线程运行时，Zed 会请求系统不要休眠。这样可以保住长回合（agent.prevent_idle_sleep）。

remote development 把语言服务器、任务与终端放到你自己的远端服务器。官方注明 v0.157 起不再经 Zed 服务器中转。

官方未见「关机后续跑 / 厂商托管后台任务」类能力（已查 Agent Panel 页）。
A local desktop app; agent threads run inside the local process. Remote hosts can be your own SSH / WSL hosts.

Bound to the client being online. While a thread runs, Zed asks the system not to sleep. This preserves long turns (agent.prevent_idle_sleep).

Remote development puts language servers, tasks and terminals on your own remote server. The official docs note since v0.157 it no longer relays through Zed servers.

No "keeps running after shutdown / vendor-hosted background task" capability found officially (Agent Panel page checked).
工具与扩展Tools & extensions: 明确支持 MCP。

官方文档 mcp.md 列明当前支持 Tools 与 Prompts 两类特性。 Discovery / Sampling / Elicitation 尚未覆盖，官方欢迎贡献。

另处理 notifications/tools/list_changed——server 运行时增删改工具时， Zed 自动重载列表，无需重启 server。

MCP 有两种接入方式：

作为 extension 安装
在 Settings → AI → MCP Servers 里添加本地 / 远程 server

官方列出的热门 extension server：Context7、GitHub、Puppeteer、Gem、 Brave Search、Framelink Figma、Resend 等。
MCP support is explicit.

The official docs mcp.md states Tools and Prompts are the two supported protocol features. Discovery / Sampling / Elicitation are not yet covered; the official docs welcome contributions.

It also handles notifications/tools/list_changed — when a server adds/removes/changes tools at runtime, Zed reloads the list automatically without restarting the server.

MCP has two integration paths:

install as an extension
add local / remote servers under Settings → AI → MCP Servers

Popular extension servers listed officially: Context7, GitHub, Puppeteer, Gem, Brave Search, Framelink Figma, Resend and more.
上下文与记忆Context & memory: Zed Agent 的上下文由 instructions + skills + MCP 组合而成（见 agents.md 的 Zed Agent 行）。

压缩策略：官方称线程接近阈值时自动 compact。默认 90% 触发（agent.auto_compact）。它把较早消息摘要后替换进模型上下文。线程显示 "Context Compacted"，可展开。

也可 /compact 手动压缩。阈值可配：百分比 / 正 token / 负剩余 token。

模型窗口太小（<80000 tokens）时自动压缩失效。官方提示用 New From Summary 新开线程。

@-mention 可加文件、目录、符号、历史线程、skills、diagnostics、分支 diff 与 URL。

另支持并行多线程（各自上下文窗口）与 Checkpoints 逐消息回滚。
Zed Agent's context is composed of instructions + skills + MCP (see the Zed Agent row in agents.md).

Compaction policy: officially, threads auto-compact when approaching the threshold. Default trigger at 90% (agent.auto_compact). It summarizes earlier messages and replaces them in the model context. The thread shows "Context Compacted", expandable.

/compact also compacts manually. The threshold is configurable: percentage / positive tokens / negative remaining tokens.

When the model window is too small (<80000 tokens) auto-compaction is disabled. The official hint is to start a new thread via New From Summary.

@-mention can add files, directories, symbols, past threads, skills, diagnostics, branch diffs and URLs.

Parallel multi-threads (each with its own context window) and Checkpoints per-message rollback are also supported.
权限与限制Permissions & limits: 工具权限由 Zed 官方配置控制（tool-permissions 页，v0.224.0+ 起由 agent.tool_permissions.default 控制）。

三档取值是 "allow" / "deny" / "confirm"，default 默认 confirm。可按工具配正则模式：always_allow / always_deny / always_confirm。

优先级：always_deny > always_confirm > always_allow > 工具 default > 全局 default。

另有硬编码安全规则（rm -rf /、rm -rf ~ 等），不可覆盖。

受管控工具含：

terminal、edit_file、write_file
delete_path、move_path、copy_path、create_directory
fetch、search_web、skill
MCP（mcp:<server>:<tool>）

另有 OS 级沙箱（sandboxing 页）：仅 Zed Agent，覆盖 terminal 与 fetch，默认限制项目外写入与出网。

权限边界跟随 agent path（agents.md）：

Zed Agent —— 用 Zed 配置的 provider 与原生工具
External Agents —— 用 ACP agent 进程自己的 auth/config
Terminal Threads —— 用原生 CLI 的 auth/config

隐私：官方称 Zed 不存 prompt 与代码上下文，遥测可关。
Tool permissions are controlled by Zed's official config (tool-permissions page; since v0.224.0 by agent.tool_permissions.default).

Three values: "allow" / "deny" / "confirm", with default set to confirm. Per-tool regex patterns can be configured: always_allow / always_deny / always_confirm.

Precedence: always_deny > always_confirm > always_allow > tool default > global default.

There are also hardcoded safety rules (rm -rf /, rm -rf ~ etc.), not overridable.

Managed tools include:

terminal, edit_file, write_file
delete_path, move_path, copy_path, create_directory
fetch, search_web, skill
MCP (mcp:<server>:<tool>)

There is also an OS-level sandbox (sandboxing page): Zed Agent only, covering terminal and fetch, restricting writes outside the project and outbound network by default.

Permission boundaries follow the agent path (agents.md):

Zed Agent — uses Zed-configured providers and native tools
External Agents — use the ACP agent process's own auth/config
Terminal Threads — use the native CLI's auth/config

Privacy: officially Zed states it does not store prompts or code context; telemetry can be turned off.
适合什么任务Best for: 适合重视编辑器性能（Rust）、解析准确度（Tree-sitter）与多人协作的场景。

也适合想在一个编辑器里混用多种 CLI Agent 的用户 —— Claude / Codex / OpenCode / Copilot / Cursor 都能挂。

这是 Zed 最独特的价值：它是 agent 的编排层，不是单一 agent。

不适合把「编辑器自带 AI」当唯一卖点去找的用户 —— Zed 的定位是编排层。

也不适合需要实测才知道大仓库表现的人 —— 索引策略官方未公开说明。
Fits scenarios that value editor performance (Rust), parsing accuracy (Tree-sitter) and multiplayer collaboration.

Also fits users who want to mix multiple CLI Agents in one editor — Claude / Codex / OpenCode / Copilot / Cursor can all be attached.

This is Zed's most unique value: it is an orchestration layer for agents, not a single agent.

Does not fit users looking for "the editor ships its own AI" as the sole selling point — Zed positions itself as the orchestration layer.

Also does not fit people who need measured runs to know large-repo behavior — the indexing strategy is not disclosed officially.

头号误解Top misconception

把 Zed 当成又一个 AI 编码工具，它实际是「三种 agent path 的编排层」，能用 Claude / Codex / OpenCode / CursorTreating Zed as yet another AI coding tool — it is an orchestration layer for three agent paths, and can use Claude / Codex / OpenCode / Cursor
以为模型是绑定的，实际有 5 条接入路径，含本地模型与复用现有订阅Assuming the model is locked in — there are five integration paths, including local models and reusing an existing subscription
选了 External Agent 后以为权限由 Zed 管，实际由那个 agent 自己的 auth/config 决定Choosing External Agent and assuming Zed governs permissions — they are determined by that agent's own auth/config

未知项清单Known unknowns

三类 agent path 在「客户端关闭后」的实际行为 Zed Agent 已知绑定客户端在线。 External / Terminal 取决于所挂 agent。

The actual behavior of the three agent paths "after the client closes" Zed Agent is known to bind to the client being online. External / Terminal depend on the attached agent.

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
docs	Zed · LLM Providers（五条模型接入路径）Zed · LLM Providers (five model-access paths)	https://zed.dev/docs/ai/llm-providers
docs	Zed · Agents（三类 agent path 对比）Zed · Agents (comparison of the three agent paths)	https://zed.dev/docs/ai/agents
docs	Zed · MCP（支持的协议特性与三种 agent path 的 MCP 行为）Zed · MCP (supported protocol features and MCP behavior across the three agent paths)	https://zed.dev/docs/ai/mcp
changelog	Zed · Releases（v1.21.0 @ 2026-09-23）Zed · Releases (v1.21.0 @ 2026-09-23)	https://github.com/zed-industries/zed/releases
repo	Zed · 仓库Zed · repository	https://github.com/zed-industries/zed
docs	Zed · 文档站Zed · docs site	https://zed.dev/docs
docs	Zed · Agent Panel（线程、checkpoints、自动压缩、@-mention、token 用量）Zed · Agent Panel (threads, checkpoints, auto-compaction, @-mention, token usage)	https://zed.dev/docs/ai/agent-panel
docs	Zed · Tool Permissions（allow/deny/confirm、正则模式、内置安全规则）Zed · Tool Permissions (allow/deny/confirm, regex patterns, built-in safety rules)	https://zed.dev/docs/ai/tool-permissions
docs	Zed · Agent Sandboxing（仅 Zed Agent、terminal/fetch、默认项目外写入与出网受限）Zed · Agent Sandboxing (Zed Agent only, terminal/fetch, out-of-project writes and egress restricted by default)	https://zed.dev/docs/ai/sandboxing
docs	Zed · Tools（内置 read/search/edit/terminal 工具清单）Zed · Tools (built-in read/search/edit/terminal tool list)	https://zed.dev/docs/ai/tools
docs	Zed · Remote Development（远端 SSH server、v0.157 起不经 Zed 中转、大目录限制）Zed · Remote Development (remote SSH server, no Zed relay since v0.157, large-directory limits)	https://zed.dev/docs/remote-development

实测记录Measured runs

本档未跑端到端编码任务；能力结论来自官方文档，编码向测量协议见 protocols/coding/_protocol.md。

实测建议：这个工具的实测重点应是混合 agent 场景。比如在 Zed 里挂 Claude Code 做 T01 任务，与直接用 Claude Code CLI 对比。要验证的是：「编排层 + 外部 agent」是否比纯 CLI 更顺手，还是只多一层壳。

This site has not completed measured runs.

The test protocol is at protocols/coding/_protocol.md.

Suggested measurement: the testing focus for this tool should be mixed-agent scenarios. E.g. attach Claude Code inside Zed to do the T01 task, compared against using the Claude Code CLI directly. What to verify: whether "orchestration layer + external agent" is handier than pure CLI, or just one extra shell.

最重要的发现：Zed 自己就用「harness」这个词The most important finding: Zed itself uses the word "harness"

官方 agents.md 把「agent path」和「harness」当成同一件事。

官方原文："An agent path is sometimes called a harness." "It is the way agentic work is started, displayed, configured, and controlled in Zed."

这句话把 AgentClash 的「Harness 章节」落到了产品文档里。 Agent 产品对比和 Harness 对比不是两件事。 AgentClash 区分它们，是因为它们落在不同的层。而 Zed 明确说「agent path 有时也叫 harness」。

The official agents.md treats "agent path" and "harness" as the same thing.

Official wording: "An agent path is sometimes called a harness." "It is the way agentic work is started, displayed, configured, and controlled in Zed."

This sentence lands AgentClash's "Harness chapter" into product documentation. Comparing Agent products and comparing Harnesses are not two different things. AgentClash distinguishes them because they sit at different layers. And Zed explicitly says "an agent path is sometimes called a harness".

三类 agent path（官方表格原文）Three agent paths (official table wording)

Agent path	运行在	用什么	适合
Zed Agent	Agent Panel + Threads 侧栏	Zed 配置的 LLM provider、原生工具、skills、instructions、MCP	想要 Zed 的原生集成
External Agents	Agent Panel + Threads 侧栏	ACP agent 进程及其自己的 auth/config	想用 Claude、Codex、OpenCode、Copilot、Cursor、Pi 等
Terminal Threads	Threads 侧栏 + 终端	原生 CLI / TUI 的 auth/config	想保留该工具的命令行体验

这张表的信息量很大：

External Agents 明确支持 Claude、Codex、OpenCode、Copilot、Cursor —— 都是本仓库收录的对象
权限由谁决定，看第二、三列 —— 选了 External Agent，auth/config 就是那个 agent 的
Zed 把自己定位成「怎么启动、显示、配置、控制 agent」的那一层

Agent path	Runs in	Uses	Fits
Zed Agent	Agent Panel + Threads sidebar	Zed-configured LLM providers, native tools, skills, instructions, MCP	wanting Zed's native integration
External Agents	Agent Panel + Threads sidebar	the ACP agent process and its own auth/config	wanting Claude, Codex, OpenCode, Copilot, Cursor, Pi etc.
Terminal Threads	Threads sidebar + terminal	the native CLI / TUI's auth/config	keeping that tool's command-line experience

This table carries a lot of information:

External Agents explicitly support Claude, Codex, OpenCode, Copilot, Cursor — all objects collected in this repository
who decides permissions — look at columns two and three — choose External Agent and auth/config belongs to that agent
Zed positions itself as the layer for "how agents are started, displayed, configured, controlled"

五条模型接入路径（这个分区独一份）Five model-access paths (unique in this track)

路径	适合	真相源
Zed-Hosted Models	想让 Zed 代计费	Account & Billing
API Access	已有 provider 额度	含 Anthropic 兼容 / OpenAI 兼容
Existing Subscription	已经付了 ChatGPT / Claude / Copilot	复用订阅
Gateway	经 OpenRouter / Bedrock / Vercel 路由	网关配置
Local Model	本地跑或自托管	本地配置

第 3 条和第 5 条是差异化：

复用已有订阅 = 边际成本可以是 0
本地模型 = 数据完全不出机器

另一条重要区分，官方明确：

官方原文："Model access paths do not configure External Agents or Terminal Threads." "External Agents and Terminal Threads usually own their own model access, auth, and configuration."

即：Zed 的模型设置管不到外部 agent。 这是很容易误解的点。

Path	Fits	Source of truth
Zed-Hosted Models	letting Zed bill on your behalf	Account & Billing
API Access	already have provider quota	includes Anthropic-compatible / OpenAI-compatible
Existing Subscription	already paying for ChatGPT / Claude / Copilot	reuse the subscription
Gateway	routing via OpenRouter / Bedrock / Vercel	gateway config
Local Model	running locally or self-hosted	local config

Paths 3 and 5 are the differentiators:

reusing an existing subscription = marginal cost can be 0
local models = data never leaves the machine

Another distinction the official docs make explicit:

Official wording: "Model access paths do not configure External Agents or Terminal Threads." "External Agents and Terminal Threads usually own their own model access, auth, and configuration."

That is: Zed's model settings do not reach external agents. This is an easy point to misunderstand.

MCP 支持的确切范围The exact scope of MCP support

项目	状态
Tools	✅ 支持
Prompts	✅ 支持
Discovery / Sampling / Elicitation	❌ 尚未覆盖（官方欢迎贡献）
`notifications/tools/list_changed`	✅ 处理——server 运行时改工具列表会自动重载，无需重启

两种接入方式：

1. 作为 extension 安装（官方列了 Context7 / GitHub / Puppeteer / Gem / Brave Search / Figma / Resend 等） 2. 在 Settings → AI → MCP Servers 里添加本地或远程 server

三种 agent path 的 MCP 行为不同：

agent path	MCP 行为
Zed Agent	直接用 Zed 配置的 MCP server
External Agents	Zed 可通过 ACP 转发已配置的 server；agent 也可读自己的原生 MCP 配置
Terminal Threads	原生 CLI/TUI 读自己的 MCP 配置

Item	Status
Tools	✅ supported
Prompts	✅ supported
Discovery / Sampling / Elicitation	❌ not yet covered (contributions welcome officially)
`notifications/tools/list_changed`	✅ handled — a server changing its tool list at runtime auto-reloads, no restart

Two integration paths:

1. install as an extension (officially listed: Context7 / GitHub / Puppeteer / Gem / Brave Search / Figma / Resend etc.) 2. add local or remote servers under Settings → AI → MCP Servers

MCP behaves differently across the three agent paths:

agent path	MCP behavior
Zed Agent	uses Zed-configured MCP servers directly
External Agents	Zed can forward configured servers via ACP; the agent can also read its own native MCP config
Terminal Threads	the native CLI/TUI reads its own MCP config

与本站其他站点的关系Relationship with this site's other sections

Zed 的 External Agents 直接涉及本仓库收录的多个对象：

Zed 里的 agent path	对应本仓库条目
External Agents 列表	Claude Code · Codex IDE · Cline · Copilot · Cursor
Terminal Threads	本仓库 agents 分区的 CLI 形态对象
MCP 扩展	tools 分区

这意味着 Zed 是本仓库的交叉点。 agents 分区两种形态（IDE 扩展 / 终端）与 tools 分区（MCP 扩展）的边界，它画得最清。

Zed's External Agents directly involve multiple objects collected in this repository:

agent path in Zed	corresponding entries in this repository
External Agents list	Claude Code · Codex IDE · Cline · Copilot · Cursor
Terminal Threads	CLI-form objects in this repository's agents section
MCP extensions	tools section

This makes Zed a crossroads of this repository. It draws the boundary between the two forms of the agents section (IDE extension / terminal) and the tools section (MCP extensions) most clearly.

采集限制（诚实说明）Collection limits (stated honestly)

本次从 zed-industries/zed 仓库的 docs/src/ai/ 逐篇读取官方文档。 八维已全部核到官方机制。含模型接入 5 路径、工具权限与 OS 级沙箱、自动压缩、MCP、多根与远端开发。

仍未核验（官方未给或需实测）：

三类 agent path 在「客户端关闭后」的实际行为

Zed Agent 已知绑定客户端在线。 External / Terminal 取决于所挂 agent 自己的配置。

This pass read the official docs one by one from docs/src/ai/ in the zed-industries/zed repository. All eight dimensions have been verified against official mechanics. Including the 5 model-access paths, tool permissions and the OS-level sandbox, auto-compaction, MCP, multi-root and remote development.

Still not verified (not given officially or needs measured runs):

the actual behavior of the three agent paths "after the client closes"

Zed Agent is known to bind to the client being online. External / Terminal depend on the attached agent's own configuration.