EasyAI

Connect the current VS Code project to any MCP-capable AI, reliably.

EasyAI creates an independent Workspace MCP connection for every local project. After you enter your user key, the extension automatically checks the account validity period, identifies the current project, completes WebCodex pairing and the Runner connection, and gives you a project MCP link you can paste straight into your AI.

Website and full tutorial:

https://mcp.oai9.com/

Official extension ID: mcp-oai9-com-EasyAI.easyai-mcp. Marketplace display name: EasyAI MCP.

Official builds target Windows x64, macOS Intel (x64), and macOS Apple silicon (arm64); VS Code 1.106.0 or newer is required. The macOS native components are ad-hoc signed and not notarized by Apple. The connection service requires a valid EasyAI user key; installing the extension does not grant free or indefinite service access.

Global Skills / MCP management

The settings page has five tabs: Account, Skills, MCP, Diagnostics, and About. Skills and MCP form a global asset library: install or add once, and every existing and future project inherits it by default — you only set per-project disable exceptions where needed. It supports importing local Skill directories with read-only SKILL.md preview, manual or JSON-based MCP configuration, searching skills.sh and the official Registry, local environment detection (Node.js / npm / Python / uv / Git), and one-click installation of npm / PyPI / GitHub Skills (installed into the extension's private directory; package scripts are never executed).

Read-only remote access to Skills. While the trusted Runner owner window of the current project is online, the project MCP link can read globally inherited Skills on demand via skills_list, skill_read, and skill_resource_read. Bodies and text resources are digest-verified and read in bounded chunks; scripts inside packages are never executed. Closing the owner window, switching projects, disabling globally, or setting a project exception blocks subsequent reads; content the AI has already received cannot be recalled. If the gateway has not enabled this channel, local management still works and the UI shows that the read-only channel is not ready.

Remote discovery and direct invocation of third-party MCP servers (stdio and HTTP/HTTPS). The trusted owner window of the current project exposes mcp_servers_list, mcp_tools_list, and mcp_tool_call over a separately authenticated channel. Remote discovery/invocation can start configured stdio servers or connect to configured Streamable HTTP endpoints on demand, without per-call or remote-start approval prompts. "Globally enabled minus project-disabled" still applies, and the AI cannot specify arbitrary programs, working directories, addresses, or install commands.

stdio MCP programs must already be installed and configured globally. command may be a program name (python, node, uv, uvx, npx, npm, or anything on PATH) or an absolute path to the program (.exe on Windows); program names are resolved to absolute paths at launch via local environment detection (including Node/Python paths manually set on the settings page) and then PATH, and npx/npm are rewritten to node plus the npx-cli.js/npm-cli.js that ship with Node.js. Script paths and options go into args. Execution never goes through a shell: command strings, relative paths, .cmd/.bat/.ps1 batch shims, and automatic installation are all rejected. Child processes inherit only location-type environment variables (system/user directories, APPDATA, etc.) and a PATH composed of the program directory plus system directories — never the host's credentials, proxies, or NODE_OPTIONS; on Windows, Python defaults to PYTHONIOENCODING=utf-8. Merely saving a configuration, opening settings, or listing servers never starts a program; tool discovery and tool calls start it on demand. The manual "start and detect tools" action keeps the local confirmation flow (the dialog shows both the configured command and the actually resolved program), but remote calls do not depend on clicking it first.

For HTTP/HTTPS MCP servers, just enter the Streamable HTTP endpoint (MCP servers exposed by other local software are typically http://127.0.0.1:port/mcp); credentials go into a Bearer token / request headers, are saved to secure storage, and are sent with every request. The URL must not contain user:password, query parameters, or a # fragment; OAuth login flows, the legacy SSE transport, redirects, and self-signed certificates are not supported. The extension only POSTs JSON-RPC and accepts JSON or event-stream responses; it does not subscribe to the server's GET event stream and does not auto-reconnect; on stop it sends DELETE to end the session. For non-local plaintext http:// addresses, the extension warns explicitly on add and on connect: request headers (including credentials) and content travel unencrypted — recommended only for trusted LANs.

Discovery and invocation validate live policy, project identity, and the tool-catalog version, with support for pagination, timeouts, and cancellation, returning native MCP content and structured results. If a tool call times out, the connection drops, or the result cannot be trusted after dispatch, MCP_CALL_OUTCOME_UNKNOWN is returned and the call is never retried automatically; the operation may already have had side effects, and stopping cannot roll them back.

"Stop" only ends the current instance; the next legitimate call may start it again. To block future calls, disable globally or set a project disable exception. Closing the panel does not stop the host; disconnecting the EasyAI account revokes the remote channel and in-flight requests, but is not the same as stopping all idle hosts or deleting third-party credentials. You can use "stop all MCP in this window". While a stop is unconfirmed, the lease is retained and never preempted automatically.

Third-party programs run with the current OS user's privileges — this is not a sandbox. Per-project process/working-directory separation is not file-access isolation. Such a program can read and write files and access the network; tool catalogs and results are forwarded through the gateway to the AI you chose and may contain paths and private data. Literal replacement of known credentials is not comprehensive DLP; object keys, protocol structures, and binary payloads are not exhaustively redacted.

Current limitations: OAuth-based MCP, the legacy SSE transport, managed install/update, the 2026 protocol, and crash recovery are not yet supported; Skill scripts are never executed. When an older extension version has not negotiated mcp-tools-v1, new calls return "unsupported" and no unknown RPCs are sent to it; the existing account, Runner, and Skills paths are unaffected. This capability does not change existing WebCodex permissions.

Highlights

One user_key can connect multiple projects.
Each project gets its own Workspace Binding, MCP URL, and WebCodex project.
Reopening the same project keeps the same MCP link.
Multiple VS Code windows on the same project share one MCP link.
At most one Runner actually runs per project at a time, preventing duplicate execution and project cross-talk.
Different projects can be online simultaneously, each using its own MCP.
After you enter the user key, the extension automatically shows expiry, remaining time, the current project, the Workspace Binding, and MCP status.
Copy "current project MCP link" or "link + recommended prompt" with one click.
A dedicated EasyAI MCP activity view.
Workspace identity registration does not directly transmit the local project's absolute path; tool output may still contain paths.

Quick start

1. Install the extension

Search @id:mcp-oai9-com-EasyAI.easyai-mcp in the VS Code Extensions view, or install the platform-specific easyai-mcp-<version>-<platform>.vsix. After installing, reload when prompted:

Developer: Reload Window

Migrating from an old preview build: first stop the EasyAI Runner in all old windows, disable or uninstall arena-platform-local.arena-runner-preview, then reload the window and install the official build. Do not keep both extensions enabled at once — they use the same command and view identifiers.

The new extension does not read the old extension's SecretStorage. Re-enter your user key, confirm workspace authorization, and connect. The device identity is regenerated, so the project MCP link may change; update the new link in your AI client. Existing accounts do not need to repurchase, and this does not automatically delete old workspace bindings on the server.

2. Open a project

Open a local project folder in VS Code and make sure the workspace is Trusted.

EasyAI only creates or queries a project MCP connection when a trusted local project exists.

3. Open EasyAI settings

Open the Command Palette and run:

EasyAI: Settings (Account, Skills & MCP)

4. Enter your user key

After you enter the user_key, the extension automatically fetches account and current-project MCP information.

This automatic lookup is preview-only:

the user key is not saved
no Runner is started
the project is not modified

Only after you click Save and auto-connect does the user key enter VS Code SecretStorage and the real connection for the current project begin.

5. Save and auto-connect

After clicking Save and auto-connect, the extension automatically:

computes a stable Workspace ID for the current project
creates or reuses the project's Workspace Binding
obtains project-specific Plugin / MCP credentials
pairs with WebCodex
starts the local Runner
links the current project
waits for the server to confirm the project is online

Once the status shows the current project is online, you can copy the MCP.

6. Hand the MCP to your AI

The settings page offers two buttons.

Copy the current project's MCP link

Copies only the project URL:

https://mcp.oai9.com/mcp/w_xxx/...

Copy link + prompt

Copies the MCP URL together with:

快速连接这个 MCP（URL），明确使用规则，熟悉可用工具，做好处理接下来一系列工作的准备。

(Roughly: "Connect to this MCP (URL), learn its usage rules, get familiar with the available tools, and be ready for the upcoming tasks.")

Recommended the first time you hand a project to an AI.

Workspace MCP rules

Reopening the same project

The MCP link stays the same.

As long as it is the same machine, the same VS Code profile / device secret, the same project path, and the same user_key — and the Workspace Binding was not deleted — reopening the project continues to use the original MCP URL.

Closing the project only takes the Runner offline temporarily; it does not delete the project's MCP identity.

Multiple VS Code windows on one project

All windows share the same Workspace ID, Workspace Binding, and MCP URL.

Opening more windows never produces extra MCP links. At most one window per project holds the Runner lease; other windows reuse the project connection state — they neither start a second Runner nor disconnect the owner window by mistake.

Different projects with the same user_key

Different projects get fully independent MCPs:

Project A -> Workspace A -> MCP A -> Runner A
Project B -> Workspace B -> MCP B -> Runner B

Both projects can be online at the same time.

Moving or renaming a project

The current version identifies a Workspace by the normalized local project path.

A moved or renamed project is treated as a new Workspace and gets a new MCP link.

Switching machines or VS Code profiles

A new Workspace identity and a new MCP link are generated by default, so one MCP URL never drifts between different physical executors.

The MCP link is a sensitive credential

EasyAI uses capability MCP URLs by default, for example:

https://mcp.oai9.com/mcp/w_xxx/<token>

The full URL itself contains the access credential.

Never put the full MCP URL in:

public chats
publicly shared screenshots
public documents
Git commits
public issues or logs
the hands of untrusted third parties

If a full MCP URL has leaked publicly, delete or rebuild the corresponding Workspace Binding to generate a new project MCP link.

user_key is equally sensitive.

Privacy design

EasyAI's workspace identity registration does not directly transmit the local project's absolute path; this is not a zero-path promise for tool catalogs, content, or results.

The extension locally combines a device secret with the normalized project path via HMAC to derive a Workspace ID from which the real path cannot be recovered.

This guarantee covers the workspace identity registration process, not all MCP traffic. Actual tool calls may read files, run commands, and return file contents, output, or embedded paths to your AI client. Evaluate the project's sensitive data before authorizing. Details ship with the package in PRIVACY.md.

The account key is stored in VS Code SecretStorage. The extension calls https://mcp.oai9.com/ for account, pairing, workspace, and activity information; the Runner connects to the configured WebCodex upstream, currently https://webcodex.oai9.com/. Tool requests and results are additionally processed by the AI/MCP client you chose. The extension keeps a bounded local activity summary; the server may retain account, connection, and execution activity records. Stopping, clearing the local key, or uninstalling the extension does not delete server-side records. For data retention and deletion requests, contact the operator via the website.

EasyAI MCP activity

The extension provides a dedicated EasyAI MCP activity view showing the current project's MCP call activity.

Open it with:

EasyAI: Open MCP Activity

No extra EasyAI panel is added to the Explorer, so your Explorer space stays clean.

Common commands

Command	Purpose
`EasyAI: Settings (Account, Skills & MCP)`	enter the key, check account validity, project connection status, copy the MCP
`EasyAI: Open MCP Activity`	open the MCP activity view
`EasyAI: Connect WebCodex Workspace`	manually trigger the current project's connection
`EasyAI: Stop Connection (stop the local Runner)`	stop the local Runner held by this window
`EasyAI: Show Status Log`	view redacted diagnostic logs

Current platform scope

Platform	Status
Windows x64	official target with completed automated acceptance for native components, installed VSIX, real Runner/WebSocket/tool round-trips, packaged CLI pairing/login, and the formal release build
Linux x64	official target with the same level of automated acceptance as Windows x64; shipped as a separate Linux x64 VSIX
macOS x64 / arm64	darwin-x64 / darwin-arm64 platform packages with completed installation, Runner/WebSocket, and packaged CLI pairing/login acceptance; native components are ad-hoc signed, without a Developer ID or Apple notarization, and lack real quarantine/Gatekeeper download acceptance
Windows arm64 / Linux arm64 / others	not officially supported at this time

The pairing/login acceptance above uses a local loopback fixture to validate the real packaged CLI's protocol, credential files, and Runner configuration generation; it does not mean the production account service itself has completed release acceptance.

EasyAI does not support remote VS Code workspaces via Remote SSH, WSL, or Dev Containers.

FAQ

No MCP shown after entering the user_key

Check that:

a local project is open
the workspace is Trusted
you are not in Remote SSH / WSL / a container
the user_key is well-formed and not expired
the MCP service is reachable

WC_PAIRING_IN_PROGRESS

A pairing flow already exists for this Workspace. Wait for it to finish or for the lease to expire; do not repeatedly create new account keys or delete database files.

WC_RUNNER_NOT_READY

The Runner has started but WebCodex has not yet confirmed the project is online. This usually resolves itself shortly; if it persists, check the network and the status log.

WC_PROVISIONING_FAILED

Server-side pairing or project linking failed. Run EasyAI: Show Status Log to see the redacted error code.

Does the MCP still work after closing VS Code?

The project's Workspace Binding and MCP URL are retained, but with the local Runner offline, actual project tools cannot execute.

Reopen the project and connect; the original MCP URL works again.

Does clearing the user_key change the project's MCP identity?

Clearing the account key does not delete the device-level Workspace identity secret, so reconnecting the same project path with the same account keeps the stable Workspace identity logic.

Security boundary

The EasyAI Runner runs with the current OS user's privileges. It is not an operating-system sandbox.

When an MCP tool can write files, run commands, or modify code, those capabilities ultimately execute in the local project environment.

Hand the project MCP link only to AI / MCP clients you trust, and make sure the current workspace is suitable for automated operation.

Website & tutorial

Full beginner tutorial, connection rules, and FAQ:

https://mcp.oai9.com/

The help link in the top-right corner of the extension's settings page opens the same site.

Development verification

npm ci
npm run typecheck
npm test
npm run package

Release packaging runs additional VSIX integrity verification to ensure the required runtime files, the MCP activity UI, the settings page, and the extension icon are all included in the final VSIX.

License & third-party components

EasyAI's original code is proprietary: installing and using official release packages is permitted, but no open-source license is granted. See LICENSE.txt inside the package for the full terms. Third-party components are used under their respective licenses; see the bundled THIRD_PARTY_NOTICES.md and runtime license files.

简体中文

English | 简体中文

让 VS Code 中的当前项目，稳定连接给支持 MCP 的 AI。

EasyAI 会为每个本地项目创建独立的 Workspace MCP 连接。输入用户密钥后，插件会自动查询账号有效期、识别当前项目、完成 WebCodex 配对与 Runner 连接，并提供可直接复制给 AI 的项目 MCP 链接。

官网与完整使用教程：

https://mcp.oai9.com/

正式扩展 ID：mcp-oai9-com-EasyAI.easyai-mcp。Marketplace 显示名称：EasyAI MCP。

正式构建面向 Windows x64、macOS Intel（x64）和 macOS Apple 芯片（arm64）；要求 VS Code 1.106.0 或更新版本。macOS 原生组件为 ad-hoc 签名，未经 Apple 公证。连接服务需要有效的 EasyAI 用户密钥；安装插件不代表获得免费或无限期的服务权限。

Skills / MCP 全局管理

设置页提供账号、Skills、MCP、诊断、关于五个页面。Skills 与 MCP 是全局资产库：安装/添加一次，所有现有和新项目默认继承，只需为个别项目设置禁用例外。支持本地 Skill 目录导入与 SKILL.md 只读预览、MCP 手动/JSON 配置、skills.sh 与官方 Registry 搜索、本机环境检测（Node.js / npm / Python / uv / Git），以及 npm / PyPI / GitHub Skill 的一键安装（安装到插件私有目录，不执行包脚本）。

Skills 远程只读读取。 当前项目的可信 Runner owner 窗口在线时，项目 MCP 链接可通过 skills_list、skill_read、skill_resource_read 按需读取全局继承的 Skills。正文与文本资源经过摘要核对、有界分块读取，不执行包内脚本。关闭 owner、项目切换、全局禁用或项目例外会阻止后续读取；不能撤回 AI 已收到的内容。网关未启用该通道时，本地管理继续可用，界面会显示只读通道未就绪。

第三方 MCP（stdio 与 HTTP/HTTPS）远程发现和直接调用。 当前项目的可信 owner 窗口通过独立认证通道提供 mcp_servers_list、mcp_tools_list、mcp_tool_call。远程发现/调用可按需启动已配置的 stdio 服务或连接已配置的 Streamable HTTP 地址，不弹出逐次调用或远程启动审批。全局启用减项目禁用仍生效，不接受 AI 指定任意程序、工作目录、地址或安装命令。

stdio 类 MCP 程序须已安装并全局配置。command 可以是程序名（python、node、uv、uvx、npx、npm 或其它 PATH 中的程序）或程序的绝对路径（Windows 为 .exe）；程序名在启动时按「本机环境检测（含设置页手动指定的 Node/Python 路径）→ PATH」解析为绝对路径，npx/npm 改写为 node + 随 Node.js 安装的 npx-cli.js/npm-cli.js。脚本路径与选项放在 args 里。始终不经 shell 执行：不接受命令串、相对路径、.cmd/.bat/.ps1 批处理垫片，也不自动安装。子进程只继承位置类环境变量（系统/用户目录、APPDATA 等）与程序目录 + 系统目录组成的 PATH，不继承宿主的凭据/代理/NODE_OPTIONS；Windows 上 Python 默认 PYTHONIOENCODING=utf-8。仅保存配置/打开设置/枚举服务不会启动程序；工具发现和工具调用会按需启动。手动“启动并检测工具”保留本地确认流程（确认框同时显示配置命令与实际解析到的程序），但远程调用不依赖先点击它。

HTTP/HTTPS 类 MCP 填写 Streamable HTTP 端点地址即可（本机其他软件提供的 MCP 通常是 http://127.0.0.1:端口/mcp），凭据放在 Bearer Token / 请求头中并保存到安全存储，随每次请求发送。地址不能带账号密码、查询参数或 # 片段；不支持 OAuth 登录流程、旧 SSE 传输、重定向和自签名证书。插件只用 POST 发送 JSON-RPC 并接收 JSON 或事件流响应，不订阅服务端 GET 事件流，不自动重连；停止时向服务端发送 DELETE 结束会话。非本机的明文 http:// 地址会在添加与连接时明确提示：请求头（含凭据）与内容在网络上不加密，仅建议用于可信局域网。

发现与调用校验实时策略、项目身份和工具目录版本，支持分页、超时和取消，返回原生 MCP 内容及结构化结果。工具调用发出后若超时、断线或结果不可信，返回 MCP_CALL_OUTCOME_UNKNOWN，不会自动重试；操作可能已经产生副作用，停止不能回滚。

“停止”只结束当前实例；下一次合法调用可以重新启动。要阻止后续调用，请全局禁用或设置项目禁用例外。 关闭面板不会停止 Host；EasyAI 账号断开会撤销远程通道/在途请求，但不等同于停止所有空闲 Host 或删除第三方凭据。可使用“停止本窗口全部 MCP”。停止未确认时保留租约，不自动抢占。

第三方程序以当前系统用户权限运行，不是沙箱。 按项目分离的进程/工作目录不构成文件访问隔离。它可以读写文件、联网；工具清单与结果会经网关发给你选择的 AI，可能含路径和私人数据。已知凭据字面量替换不是全面 DLP；对象键、协议结构和二进制载荷不会被全面脱敏。

**当前限制：**尚不支持 OAuth 类 MCP、旧 SSE 传输、托管安装/更新、2026 协议和崩溃恢复；不执行 Skill 脚本。旧版本插件未协商 mcp-tools-v1 时新调用返回不支持，不向其发送未知 RPC；原有账号、Runner 与 Skills 链路保留。此能力不改变原有 WebCodex 权限。

主要功能

一个 user_key 可以连接多个项目。
每个项目拥有独立的 Workspace Binding、MCP URL 和 WebCodex project。
同一项目关闭后重新打开，MCP 链接保持不变。
同一项目同时打开多个 VS Code 窗口时，所有窗口复用同一个 MCP 链接。
同一项目同一时间最多只有一个 Runner 实际运行，避免重复执行和项目串线。
不同项目可以同时在线，并行使用各自的 MCP。
输入用户密钥后自动查询到期时间、剩余时间、当前项目、Workspace Binding 与 MCP 状态。
支持复制“当前项目 MCP 链接”和“链接 + 推荐提示词”。
提供独立的 EasyAI MCP 活动视图。
工作区身份注册不直接发送本地项目绝对路径；工具输出可能包含路径。

快速开始

1. 安装插件

在 VS Code 扩展页搜索 @id:mcp-oai9-com-EasyAI.easyai-mcp 安装，也可使用对应平台的 easyai-mcp-<版本>-<平台>.vsix。安装后按提示重载窗口：

Developer: Reload Window

**从旧预览版迁移：**先在所有旧窗口停止 EasyAI Runner，停用或卸载 arena-platform-local.arena-runner-preview，再重载窗口并安装正式版。不要同时启用新旧两个扩展，它们使用相同的命令和视图标识。

新扩展不会自动读取旧扩展的 SecretStorage。请重新录入用户密钥，确认工作区授权并连接。设备身份会重新生成，项目 MCP 链接可能改变；请把新链接更新到 AI 客户端。旧账号无需因此重新购买，本操作也不会自动删除服务端的旧工作区绑定。

2. 打开项目

在 VS Code 中打开一个本地项目文件夹，并确认当前 Workspace 是 Trusted。

EasyAI 只有在存在一个可信的本地项目时，才会创建或查询项目 MCP 连接。

3. 打开 EasyAI 设置

打开命令面板，运行：

EasyAI: 设置（账号、Skills 与 MCP）

4. 输入用户密钥

输入 user_key 后，插件会自动查询账号与当前项目 MCP 信息。

自动查询只用于预览：

不保存用户密钥
不启动 Runner
不修改项目

只有点击 保存并自动连接 后，用户密钥才会进入 VS Code SecretStorage，并开始正式连接当前项目。

5. 保存并自动连接

点击 保存并自动连接 后，插件会自动完成：

计算当前项目稳定的 Workspace ID
创建或复用项目 Workspace Binding
获取项目专属 Plugin / MCP 凭据
配对 WebCodex
启动本地 Runner
Link 当前项目
等待服务端确认项目上线

当状态显示当前项目已在线后，即可复制 MCP。

6. 复制 MCP 给 AI

设置页提供两个按钮。

复制当前项目 MCP 链接

只复制当前项目 URL：

https://mcp.oai9.com/mcp/w_xxx/...

复制链接和提示词

会复制 MCP URL，并附带：

快速连接这个 MCP（URL），明确使用规则，熟悉可用工具，做好处理接下来一系列工作的准备。

第一次把项目交给 AI 时，推荐使用这个方式。

Workspace MCP 规则

同一个项目关闭后再打开

MCP 链接保持不变。

只要仍然是同一台机器、同一个 VS Code Profile / device secret、同一个项目路径和同一个 user_key，并且 Workspace Binding 没有被主动删除，重新打开项目后会继续使用原 MCP URL。

关闭项目只会让 Runner 暂时离线，不会删除项目 MCP 身份。

同一个项目打开多个 VS Code 窗口

所有窗口使用同一个 Workspace ID、Workspace Binding 和 MCP URL。

不会因为打开多个窗口而生成多条 MCP。同一项目最多一个窗口持有 Runner lease；其它窗口复用项目连接状态，不会启动第二个 Runner，也不会误断开 owner 窗口的连接。

不同项目使用同一个 user_key

不同项目会得到完全独立的 MCP：

Project A -> Workspace A -> MCP A -> Runner A
Project B -> Workspace B -> MCP B -> Runner B

两个项目可以同时在线。

项目移动或重命名

当前版本按本机规范化项目路径识别 Workspace。

如果项目移动或重命名，会被视为新的 Workspace，并获得新的 MCP 链接。

换电脑或 VS Code Profile

默认生成新的 Workspace 身份和新的 MCP 链接，避免同一个 MCP URL 在不同实际执行机器之间漂移。

MCP 链接是敏感凭据

EasyAI 默认使用 capability MCP URL，例如：

https://mcp.oai9.com/mcp/w_xxx/<token>

完整 URL 本身就包含访问凭据。

请不要把完整 MCP URL：

发到公开聊天
截图公开分享
写入公开文档
提交到 Git
放入公开 issue 或日志
发给不可信第三方

如果完整 MCP URL 已公开泄露，应删除或重建对应 Workspace Binding，生成新的项目 MCP 链接。

user_key 同样属于敏感凭据。

隐私设计

EasyAI 的工作区身份注册不直接发送本地项目绝对路径；这不是对工具清单、内容或结果的零路径承诺。

插件会在本地使用 device secret 与规范化项目路径，通过 HMAC 生成不可反推出真实路径的 Workspace ID。

上述限制指工作区身份注册过程，不是对所有 MCP 流量的承诺。实际工具调用可能读取文件、执行命令，并向接入的 AI 客户端返回文件内容、输出或其中包含的路径。请在授权前评估项目敏感信息。详细说明随安装包附于 PRIVACY.md。

账号密钥保存在 VS Code SecretStorage 中。插件向 https://mcp.oai9.com/ 请求账号、配对、工作区和活动信息；Runner 连接服务配置的 WebCodex 上游，当前为 https://webcodex.oai9.com/。工具请求与结果还会交给你选择的 AI/MCP 客户端处理。插件保存有界的本地活动摘要；服务端可能保留账号、连接与执行活动记录。停止、清除本地密钥或卸载插件不等于删除服务端记录。数据保留与删除请求请通过官网联系运营方。

EasyAI MCP 活动

插件提供独立的 EasyAI MCP 活动 视图，用于查看当前项目的 MCP 调用活动。

通过命令：

EasyAI: 打开 MCP 活动

即可打开。

资源管理器中不会额外创建 EasyAI 折叠面板，避免占用 Explorer 空间。

常用命令

命令	作用
`EasyAI: 设置（账号、Skills 与 MCP）`	输入密钥、查看账号有效期、项目连接状态、复制 MCP
`EasyAI: 打开 MCP 活动`	打开 MCP 活动视图
`EasyAI: 连接 WebCodex 工作区`	手动触发当前项目连接
`EasyAI: 停止连接（停止本地 Runner）`	停止当前窗口持有的本地 Runner
`EasyAI: 查看状态日志`	查看脱敏诊断日志

当前平台范围

平台	状态
Windows x64	正式目标已完成 native、Installed VSIX、真实 Runner/WebSocket/tool round-trip、packaged CLI pairing/login 与 formal release build 自动验收
Linux x64	正式目标已完成与 Windows x64 同等级自动验收；使用单独的 Linux x64 VSIX 发布包
macOS x64 / arm64	提供 darwin-x64 / darwin-arm64 平台包，已完成安装、Runner/WebSocket 与 packaged CLI pairing/login 验收；原生组件为 ad-hoc 签名，无 Developer ID、未经 Apple 公证，缺少真实 quarantine/Gatekeeper 下载验收
Windows arm64 / Linux arm64 / 其他	当前未作为正式支持平台

上述 pairing/login 自动验收使用本机 loopback fixture 验证真实 packaged CLI 的协议、凭据文件与 Runner 配置生成，不代表生产账号服务本身已经完成发行验收。

EasyAI 不支持 Remote SSH、WSL 或 Dev Container 形式的远程 VS Code Workspace。

常见问题

输入 user_key 后没有显示 MCP

确认：

已经打开一个本地项目
Workspace 已 Trusted
当前不是 Remote SSH / WSL / Container
user_key 格式正确且没有过期
MCP Service 可以访问

显示 WC_PAIRING_IN_PROGRESS

同一个 Workspace 已经存在配对流程。等待当前配对完成或租约到期，不要反复创建新账号 Key 或删除数据库文件。

显示 WC_RUNNER_NOT_READY

Runner 已启动，但 WebCodex 尚未确认项目上线。通常等待短时间后会自动恢复；如果持续出现，请检查网络和状态日志。

显示 WC_PROVISIONING_FAILED

服务端配对或项目 Link 失败。运行 EasyAI: 查看状态日志 查看脱敏错误码。

关闭 VS Code 后 MCP 还能用吗？

项目 Workspace Binding 和 MCP URL 会保留，但本地 Runner 离线后，实际项目工具无法执行。

重新打开项目并连接后，原 MCP URL 恢复工作。

清除 user_key 会不会改变项目 MCP 身份？

清除账号密钥不会删除设备级 Workspace identity secret，因此重新使用同一账号连接同一路径项目时，可以保持稳定的 Workspace 身份逻辑。

安全边界

EasyAI Runner 以当前操作系统用户权限运行，它不是操作系统沙箱。

当 MCP 工具具有写文件、执行命令或修改代码能力时，这些能力最终会在本机项目环境执行。

请只把项目 MCP 链接交给你信任的 AI / MCP 客户端，并确认当前 Workspace 适合被自动化操作。

官网与教程

完整新手教程、连接规则和常见问题：

https://mcp.oai9.com/

插件设置页右上角的 不会使用点我 也会打开这个页面。

开发验证

npm ci
npm run typecheck
npm test
npm run package

发布包会执行额外的 VSIX 完整性验证，确保必需运行时文件、MCP 活动 UI、设置页与扩展图标都包含在最终 VSIX 中。

许可与第三方组件

EasyAI 原创代码采用专有许可，允许安装和使用官方发布包，不授予开放源代码许可。完整条款见安装包内的 LICENSE.txt。第三方组件依各自许可证使用，参见随包的 THIRD_PARTY_NOTICES.md 和运行时许可证文件。