diff --git a/README.md b/README.md
index c5e1b4c..cac2267 100644
--- a/README.md
+++ b/README.md
@@ -159,6 +159,41 @@ npm run dev
 - API文档(默认): http://localhost:10392/docs
 - 也可在应用内“设置 -> 后端端口”修改（支持“恢复默认”一键回到 10392）：网页端会尝试重启本机后端到新端口并刷新（并写入 `output/runtime_settings.json`，开发模式下也会写入项目根目录 `.env` 供 `uv run` 下次启动使用）；桌面端会重启内置后端并刷新
 
+## MCP 服务
+
+后端提供 MCP JSON-RPC over HTTP 服务，默认只监听 `127.0.0.1`。手机接入局域网时，在应用内打开 **设置 -> MCP 接入 -> 允许手机局域网接入 MCP**，后端会切换为监听 `0.0.0.0` 并重启。
+
+MCP 入口需要 token 鉴权。设置页提供 **MCP Token**、**AI 接入提示词** 和 **Skill Markdown** 三个独立复制区；token 可一键复制或重置，重置后旧 token 立即失效。手机端或外部 AI 客户端访问 `/mcp`、`/mcp/skill/bundle`、`/mcp/skill` 时，都应带上 `Authorization: Bearer <MCP_TOKEN>`。兼容客户端也可以使用 `X-MCP-Token` 请求头或 `?token=` 查询参数，但推荐使用 Bearer token。
+
+通用客户端可以通过 `GET /mcp/skill/bundle` 读取同一份 bundle，通过 `GET /mcp/skill` 读取 markdown 版本，这两个 skill 入口同样需要 MCP token。
+
+工具调用成功时，客户端优先读取 `result.structuredContent`，`content[0].text` 只是给通用 MCP 客户端展示的 JSON 文本副本。业务未就绪时仍可能返回 JSON-RPC success，但 `result.isError=true`；协议错误或参数错误则返回 JSON-RPC `error`。
+
+工具按包分层：
+
+- `wechat.core`: 状态、工具目录、账号列表、账号信息
+- `wechat.setup`: 密钥读取、数据库解密、已解密目录导入、媒体密钥保存、批量媒体处理入口
+- `wechat.system`: 健康检查、后端端口、日志路径、大图辅助插件状态等系统能力
+- `wechat.mobile`: 面向手机和外部代理的聚合入口，默认返回小结果和下一步建议
+- `wechat.contacts`: 联系人列表、模糊解析、联系人导出
+- `wechat.chat`: 会话、消息、搜索、发送者筛选、上下文、锚点、合并转发/AppMsg 解析、统计
+- `wechat.moments`: 朋友圈时间线、用户、同步、图片/视频/文章封面 URL
+- `wechat.media`: 聊天/朋友圈图片、视频、表情、头像、语音文件 URL、远程图片代理与资源辅助
+- `wechat.biz`: 公众号/服务号与微信支付记录
+- `wechat.analytics`: 年度总结与卡片懒加载
+- `wechat.export`: 聊天、朋友圈、账号归档导出任务、下载 URL 与进度 SSE URL
+- `wechat.admin`: 微信检测、索引、实时同步等管理能力
+- `wechat.editing`: 消息编辑、修复、恢复与审计
+
+媒体、视频、SSE 进度和 ZIP 导出不会直接塞进 MCP JSON 响应；相关工具返回可访问 URL、`streamUrl`、任务状态或资源参数。
+
+配套 skill 可通过 HTTP 加载，访问时需要带 MCP token：
+
+- JSON bundle: `http://<电脑局域网IP>:10392/mcp/skill/bundle`
+- Markdown bundle: `http://<电脑局域网IP>:10392/mcp/skill`
+
+手机端或外部 AI 客户端应先拉取 skill bundle，将 `bundleText` 注入模型上下文，再按 `initialize`、`tools/list`、`tools/call` 使用 MCP。设置页中的“AI 接入提示词”会包含 endpoint 和 Bearer token，可直接复制给客户端作为接入指令。
+
 ## 打包为 EXE（Windows 桌面端）
 
 本项目提供基于 Electron 的桌面端安装包（NSIS `Setup.exe`）。
diff --git a/skills/wechat-mcp-copilot/SKILL.md b/skills/wechat-mcp-copilot/SKILL.md
new file mode 100644
index 0000000..bbfc8ef
--- /dev/null
+++ b/skills/wechat-mcp-copilot/SKILL.md
@@ -0,0 +1,32 @@
+---
+name: wechat-mcp-copilot
+version: "1.0.0"
+description: Use WeChatDataAnalysis MCP to inspect local WeChat accounts, contacts, sessions, messages, Moments, media, exports, and analytics through a small routed playbook. Trigger when the user asks to search, summarize, export, diagnose, or reason over local WeChat data.
+---
+
+# WeChat MCP Copilot
+
+Use WeChatDataAnalysis MCP like an investigator: start broad, resolve fuzzy targets, then fetch only the context needed to answer.
+
+## Core Rules
+
+1. Start with `references/routing.md`.
+2. Load only one domain reference after routing.
+3. Load `references/pagination-budget.md` before broad searches, exports, or multi-page scans.
+4. Use `references/failure-recovery.md` when MCP, database readiness, or empty results are unclear.
+5. For phone, ScreenMemo, or external MCP clients, prefer `wechat.mobile.*` facade tools before low-level tools.
+6. Do not load the full tool catalog unless the user asks about available tools.
+
+## References
+
+- `references/routing.md`: first-hop intent routing.
+- `references/mobile.md`: phone-friendly facade tools and compact response rules.
+- `references/setup-system.md`: setup, keys, decrypt, import, health, and system operations.
+- `references/target-resolution.md`: fuzzy contact/session resolution.
+- `references/chats.md`: chat sessions, messages, search, and context.
+- `references/moments.md`: Moments timeline, posters, likes, comments, media.
+- `references/media.md`: images, videos, emoji, files, voice resources without transcription.
+- `references/export.md`: chat, Moments, and account archive export jobs.
+- `references/analytics.md`: wrapped cards, counts, rankings, and aggregate analysis.
+- `references/pagination-budget.md`: limits, cursors, result clipping, stopping rules.
+- `references/failure-recovery.md`: empty result, not-ready database, ambiguous targets, retries.
diff --git a/skills/wechat-mcp-copilot/references/analytics.md b/skills/wechat-mcp-copilot/references/analytics.md
new file mode 100644
index 0000000..1d46163
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/analytics.md
@@ -0,0 +1,19 @@
+# Analytics
+
+Use this for annual summaries, rankings, counts, and aggregate questions.
+
+## Tools
+
+- `wechat.analytics.get_wrapped_meta`
+- `wechat.analytics.get_wrapped_card`
+- `wechat.analytics.get_wrapped_annual`
+- `wechat.chat.get_daily_message_counts`
+- `wechat.biz.get_pay_records`
+
+## Rules
+
+- Prefer `get_wrapped_meta` then `get_wrapped_card` for mobile or constrained contexts.
+- Use `get_wrapped_annual` only when the user needs the whole annual dataset.
+- For broad statistics, prefer aggregate tools or targeted searches over full message pagination.
+- Always state the account, time range, and metric basis when answering.
+
diff --git a/skills/wechat-mcp-copilot/references/chats.md b/skills/wechat-mcp-copilot/references/chats.md
new file mode 100644
index 0000000..9143bd9
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/chats.md
@@ -0,0 +1,28 @@
+# Chats
+
+Use this for chat sessions, message search, and message context.
+
+## Flow
+
+1. Resolve fuzzy target with `wechat.chat.resolve_session`.
+2. For recent messages, call `wechat.chat.get_messages` with a small `limit`.
+3. For keywords, call `wechat.chat.search_messages`.
+4. Use `wechat.chat.list_search_senders` when the user needs sender facets for a broad search.
+5. For a hit that needs context, call `wechat.chat.get_message_around`.
+6. For merged-forward chat history or AppMsg cards that only expose `server_id`, call `wechat.chat.resolve_chat_history` or `wechat.chat.resolve_app_message`.
+7. Use `wechat.chat.get_message_raw` only for debugging or missing structured fields.
+
+## Useful Tools
+
+- `wechat.chat.list_sessions`
+- `wechat.chat.resolve_session`
+- `wechat.chat.get_messages`
+- `wechat.chat.search_messages`
+- `wechat.chat.list_search_senders`
+- `wechat.chat.get_message_around`
+- `wechat.chat.get_message_anchor`
+- `wechat.chat.get_daily_message_counts`
+- `wechat.chat.resolve_chat_history`
+- `wechat.chat.resolve_app_message`
+
+Do not scan full histories by pagination when an aggregate or search tool can answer.
diff --git a/skills/wechat-mcp-copilot/references/export.md b/skills/wechat-mcp-copilot/references/export.md
new file mode 100644
index 0000000..1d61466
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/export.md
@@ -0,0 +1,29 @@
+# Export
+
+Use export only when the user asks for an artifact.
+
+## Chat Export
+
+1. Resolve target session if `scope=selected`.
+2. Confirm time range, format, media options, and output directory when needed.
+3. Preview targets with `wechat.export.preview_chat_targets`.
+4. Create job with `wechat.export.create_chat_export`.
+5. Poll `wechat.export.get_chat_export`.
+6. Return `wechat.export.get_chat_export_download` when ready.
+7. Use `wechat.export.get_chat_export_events_url` when the client can consume SSE progress.
+
+## Moments Export
+
+Use `wechat.export.create_moments_export`, `wechat.export.get_moments_export`, `wechat.export.get_moments_export_download`, and `wechat.export.get_moments_export_events_url`.
+
+## Account Archive
+
+Use `wechat.export.create_account_archive`, `wechat.export.get_account_archive`, `wechat.export.get_account_archive_download`, and `wechat.export.cancel_account_archive`.
+
+## Contacts Export
+
+Use `wechat.contacts.export_contacts` only when the user asks for a contacts file. It writes JSON or CSV to a local output directory.
+
+Do not silently export all history and all media unless the user explicitly asked for that scope.
+
+For phone clients, prefer `wechat.mobile.export_job` unless exact low-level export options are required.
diff --git a/skills/wechat-mcp-copilot/references/failure-recovery.md b/skills/wechat-mcp-copilot/references/failure-recovery.md
new file mode 100644
index 0000000..b702d1d
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/failure-recovery.md
@@ -0,0 +1,21 @@
+# Failure Recovery
+
+Use this when MCP status, DB readiness, or results are suspicious.
+
+## Checks
+
+1. Phone clients: `wechat.mobile.get_overview`
+2. `wechat.core.get_status`
+3. `wechat.core.list_accounts`
+4. `wechat.core.get_account_info`
+5. Search index status with `wechat.chat.get_search_index_status` when message search fails.
+6. Moments availability by checking account info and `wechat.moments.list_users`.
+7. Setup readiness: load `setup-system.md` for keys, decrypt, import, health, or media-key problems.
+
+## Empty Results
+
+- Do not conclude "no data" after one failed query.
+- Try contact/session resolution with a simpler keyword.
+- Try session search before global message search when a target is known.
+- For Moments, resolve poster identity before timeline filtering.
+- If setup is not ready, stop content tools and explain the missing readiness condition.
diff --git a/skills/wechat-mcp-copilot/references/media.md b/skills/wechat-mcp-copilot/references/media.md
new file mode 100644
index 0000000..b136682
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/media.md
@@ -0,0 +1,30 @@
+# Media
+
+Use this for image, video, emoji, file, link, and voice resources.
+
+## Tools
+
+- `wechat.media.get_avatar_url`
+- `wechat.media.get_chat_image_url`
+- `wechat.media.get_chat_emoji_url`
+- `wechat.media.get_chat_video_thumb_url`
+- `wechat.media.get_chat_video_url`
+- `wechat.media.get_chat_voice_url`
+- `wechat.media.get_decrypted_resource_url`
+- `wechat.media.get_proxy_image_url`
+- `wechat.media.get_favicon_url`
+- `wechat.media.open_chat_media_folder`
+- `wechat.biz.get_proxy_image_url`
+- `wechat.moments.get_media_url`
+- `wechat.moments.get_article_thumb_url`
+- `wechat.moments.get_remote_video_url`
+- `wechat.moments.get_local_video_url`
+
+## Rules
+
+- Media tools return URLs or resource metadata; they do not inline large binary payloads.
+- Voice resources are files only. Do not transcribe voice messages.
+- For phone clients, prefer `wechat.mobile.get_media_links` first.
+- `open_chat_media_folder` is a desktop-host action; do not use it for phone-only flows.
+- Locate the message first, then fetch media URL by message fields such as `server_id`, `username`, `md5`, or returned media references.
+- For Moments, prefer local media URL fields from timeline records. Use remote video/article helpers only when the timeline record has a remote URL or article URL.
diff --git a/skills/wechat-mcp-copilot/references/mobile.md b/skills/wechat-mcp-copilot/references/mobile.md
new file mode 100644
index 0000000..d34e1cd
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/mobile.md
@@ -0,0 +1,30 @@
+# Mobile Facade
+
+Use this for phone, ScreenMemo, and external MCP clients unless the user needs a low-level operation.
+
+## Default Tools
+
+- `wechat.mobile.get_overview`: first call after initialize. Returns readiness, accounts, health, and suggested next tools.
+- `wechat.mobile.get_home_snapshot`: small account/session/Moments snapshot for a home screen.
+- `wechat.mobile.resolve_target`: resolve fuzzy people, groups, sessions, Moments users, or official accounts.
+- `wechat.mobile.search_chat`: message search with optional tiny context windows.
+- `wechat.mobile.get_chat_context`: recent, day, or around-anchor chat context.
+- `wechat.mobile.get_session_bundle`: session metadata plus a page of messages.
+- `wechat.mobile.search_moments`: compact Moments search.
+- `wechat.mobile.get_media_links`: URL-only media lookup.
+- `wechat.mobile.get_analytics`: compact analytics by metric.
+- `wechat.mobile.export_job`: preview/create/status/download/cancel export jobs.
+
+## Budget Rules
+
+- Keep `limit` at 10-20 for first calls.
+- Use `offset` or returned cursor fields for paging.
+- Do not call full annual analytics by default; use `metric=digest` or a single card.
+- Do not fetch binary media through MCP. Use returned URLs in the app.
+- Use low-level tools only for debugging, editing, raw fields, unusual media, or exact export control.
+
+## Recovery
+
+- If `ready=false`, load `setup-system.md`.
+- If target resolution is ambiguous, ask for one clarifying clue or show top candidates.
+- If search returns nothing, try `resolve_target` and then `get_chat_context` before declaring no data.
diff --git a/skills/wechat-mcp-copilot/references/moments.md b/skills/wechat-mcp-copilot/references/moments.md
new file mode 100644
index 0000000..85cbaf1
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/moments.md
@@ -0,0 +1,18 @@
+# Moments
+
+Use this for 朋友圈, posts, likes, comments, shared links, and Moments media.
+
+## Flow
+
+1. If the clue is a person, resolve with `wechat.contacts.resolve_contact`.
+2. Use `wechat.moments.list_users` when you need poster candidates.
+3. Use `wechat.moments.list_timeline` or `wechat.moments.search_moments`.
+4. Use `wechat.moments.get_media_url` only when the user needs a media resource.
+
+## Rules
+
+- Person names must be resolved to username before filtering timeline by `usernames`.
+- Keyword search is for post content/topic, not poster identity.
+- Do not request raw XML by default.
+- Use `wechat.moments.sync_latest` only when the user explicitly wants fresh local sync or status indicates data is stale.
+
diff --git a/skills/wechat-mcp-copilot/references/pagination-budget.md b/skills/wechat-mcp-copilot/references/pagination-budget.md
new file mode 100644
index 0000000..9e5ec65
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/pagination-budget.md
@@ -0,0 +1,20 @@
+# Pagination And Budget
+
+Default limits:
+
+- Contact/session candidates: 10
+- Recent messages: 20
+- Message search: 20
+- Moments timeline: 10
+- Media references: 20
+- Ranking/analytics rows: 20
+
+Hard rules:
+
+- Keep single message/Moments pages at or below 50 unless user asks for more.
+- Stop paging when enough evidence exists.
+- Stop after two empty or low-value pages.
+- Do not cross 500 raw messages without user confirmation.
+- List responses should use ids, names, timestamps, preview, and evidence.
+- Fetch full details only after a candidate or hit is selected.
+
diff --git a/skills/wechat-mcp-copilot/references/routing.md b/skills/wechat-mcp-copilot/references/routing.md
new file mode 100644
index 0000000..55116c0
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/routing.md
@@ -0,0 +1,24 @@
+# Routing
+
+Use this first for every WeChatDataAnalysis MCP task.
+
+## First Calls
+
+- Phone, ScreenMemo, mobile app, or compact external client: load `mobile.md` and start with `wechat.mobile.get_overview`.
+- Status, readiness, "why can't I find anything": `wechat.core.get_status`, or `wechat.mobile.get_overview` for phone clients.
+- Available tools or packages: `wechat.core.list_tools`.
+- Account selection: `wechat.core.list_accounts`, then `wechat.core.get_account_info`.
+- Key/decrypt/import/backend health problems: load `setup-system.md`.
+- Fuzzy person/group/official account: load `target-resolution.md`.
+- Chat content, recent messages, keyword search: load `chats.md`.
+- Moments / 朋友圈 / likes / comments / post media: load `moments.md`.
+- Images, videos, emoji, files, voice resources: load `media.md`.
+- Export requests: load `export.md`.
+- Rankings, yearly summary, activity stats: load `analytics.md`.
+- Empty results or setup errors: load `failure-recovery.md`.
+
+## Mixed Intent
+
+Resolve the target first, then load only the main domain reference. Do not load chats, moments, media, export, and analytics together unless the user explicitly asks for a broad audit.
+
+For phone clients, keep using `mobile.md` until the user needs a low-level fallback such as editing, raw fields, special media URL construction, or exact export options.
diff --git a/skills/wechat-mcp-copilot/references/setup-system.md b/skills/wechat-mcp-copilot/references/setup-system.md
new file mode 100644
index 0000000..65b3c56
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/setup-system.md
@@ -0,0 +1,41 @@
+# Setup And System
+
+Use this when the database is not ready, keys are needed, decrypted data must be imported, or the backend itself needs inspection.
+
+## Setup Tools
+
+- `wechat.setup.get_saved_keys`: read saved DB/media keys for an account or wxid directory.
+- `wechat.setup.get_database_key`: desktop workflow to extract the DB key from local WeChat.
+- `wechat.setup.get_image_key`: fetch and save image AES/XOR keys.
+- `wechat.setup.decrypt_databases`: decrypt databases from `db_storage_path` and a DB key.
+- `wechat.setup.get_decrypt_stream_url`: SSE URL for decrypt progress.
+- `wechat.setup.preview_import_decrypted`: validate an already-decrypted account directory.
+- `wechat.setup.get_import_decrypted_stream_url`: SSE URL for import progress.
+- `wechat.setup.cancel_import_decrypted`: cancel an import job by `job_id`.
+- `wechat.setup.save_media_keys`: save media XOR/AES keys.
+- `wechat.setup.decrypt_all_media`: decrypt all local `.dat` image resources.
+- `wechat.setup.get_decrypt_all_media_stream_url`: SSE URL for bulk media decrypt progress.
+- `wechat.setup.get_download_all_emojis_stream_url`: SSE URL for bulk emoji download progress.
+
+## System Tools
+
+- `wechat.system.health_check`
+- `wechat.system.api_root`
+- `wechat.system.get_backend_log_file`
+- `wechat.system.open_backend_log_file`
+- `wechat.system.get_backend_port`
+- `wechat.system.set_backend_port_setting`
+- `wechat.system.set_backend_port_and_restart`
+- `wechat.system.get_img_helper_status`
+- `wechat.system.toggle_img_helper`
+- `wechat.system.pick_directory`
+- `wechat.system.log_frontend_server_error`
+
+## Rules
+
+- Stream tools return `streamUrl`; the client consumes SSE outside the MCP JSON response.
+- `set_backend_port_setting` persists the setting and may require backend restart by the user/client flow.
+- `set_backend_port_and_restart` changes the port through the desktop restart flow and will disrupt the current backend connection.
+- `open_backend_log_file` and `pick_directory` are desktop-host GUI actions; do not use them for phone-only flows.
+- DB key extraction and image helper toggling depend on the local desktop WeChat state.
+- Import/decrypt/media bulk operations write local files; summarize expected impact before running them.
diff --git a/skills/wechat-mcp-copilot/references/target-resolution.md b/skills/wechat-mcp-copilot/references/target-resolution.md
new file mode 100644
index 0000000..54c2818
--- /dev/null
+++ b/skills/wechat-mcp-copilot/references/target-resolution.md
@@ -0,0 +1,21 @@
+# Target Resolution
+
+Use target resolution whenever the user gives a loose name, nickname, group name, remark, official account name, or "the person/group who...".
+
+## Tools
+
+- Contacts: `wechat.contacts.resolve_contact`
+- Sessions: `wechat.chat.resolve_session`
+- Fallback list: `wechat.contacts.list_contacts`, `wechat.chat.list_sessions`
+
+## Rules
+
+- Prefer exact username/wxid match, then remark, nickname, display name, alias, recent session evidence.
+- If a chat task mentions a person, resolve both contact and session when needed.
+- If several candidates remain plausible, inspect recent session previews before choosing.
+- If ambiguity survives after reasonable evidence, ask a short clarification.
+
+## Evidence Fields
+
+Use username/session id, display name, remark/nickname/alias, session type, recent timestamp, recent preview, and confidence.
+