Personal OS MCP
MCP server for Personal OS API that enables managing tasks, notes, projects, collections, reviews, and more through typed tools over stdio.
README
Personal OS MCP
Independent Model Context Protocol server for the Personal OS AI API. It exposes practical, typed MCP tools over stdio and performs every real operation through the existing HTTP API.
MCP client → personal-os-mcp → Personal OS /api/v1/ai/* → Personal OS database
The server never connects to MySQL, imports Laravel internals, accepts a User ID, or reproduces Personal OS ownership and lifecycle logic. Personal OS remains authoritative.
Technology
- Node.js 20 or newer
- Node.js 22.7.5 or newer when running the current MCP Inspector
- TypeScript with strict checking
- Official
@modelcontextprotocol/sdk1.30 - Zod schemas
- Native Node
fetch - Vitest, ESLint, and Prettier
- Local
stdiotransport
Setup
Personal OS must be available locally, normally at http://localhost:8080.
cd /home/lika/projects/personal-os-mcp
npm install
cp .env.example .env
Create a dedicated User-bound AI token from the Personal OS project:
cd /home/lika/projects/personal-os
docker compose exec app php artisan ai:token user@example.com --name=chatgpt
Use the actual account email interactively. The token is displayed once. Store it only in the local ignored .env file or a secret manager:
PERSONAL_OS_BASE_URL=http://localhost:8080
PERSONAL_OS_AI_TOKEN=replace_with_the_once_displayed_token
PERSONAL_OS_MCP_TIMEOUT_SECONDS=15
Never commit, log, paste into documentation, or place a real token in .env.example.
Commands
npm run dev
npm run typecheck
npm run build
npm start
npm test
npm run lint
npm run format:check
Development mode runs TypeScript directly. npm run build writes ESM output to dist; npm start runs the built stdio server.
Inspect the built server with the official MCP Inspector:
npm run build
npx @modelcontextprotocol/inspector node dist/index.js
The server validates required environment variables at startup. Protocol messages use stdout; configuration failures use stderr without printing the token or environment.
Tools
System
personal_os_healthpersonal_os_capabilities
Tasks and Planning
personal_os_list_taskspersonal_os_get_taskpersonal_os_create_taskpersonal_os_update_taskpersonal_os_bulk_update_taskspersonal_os_complete_taskpersonal_os_reopen_taskpersonal_os_archive_taskpersonal_os_unarchive_taskpersonal_os_move_task_to_trashpersonal_os_restore_taskpersonal_os_update_task_planningpersonal_os_list_planning_taskspersonal_os_get_task_planning
Tags
personal_os_list_tagspersonal_os_get_tagpersonal_os_create_tagpersonal_os_update_tagpersonal_os_archive_tagpersonal_os_restore_tagpersonal_os_delete_tag
Containers / Lists
personal_os_list_containerspersonal_os_get_containerpersonal_os_create_containerpersonal_os_update_containerpersonal_os_archive_containerpersonal_os_restore_containerpersonal_os_delete_container
Projects
personal_os_list_projectspersonal_os_get_projectpersonal_os_create_projectpersonal_os_update_projectpersonal_os_archive_projectpersonal_os_restore_projectpersonal_os_delete_project
Collections
personal_os_list_collectionspersonal_os_get_collectionpersonal_os_create_collectionpersonal_os_update_collectionpersonal_os_archive_collectionpersonal_os_restore_collectionpersonal_os_delete_collection
A Collection is a root-level Task Container. There is no collection-item model or item mutation tool.
Notes
personal_os_list_notespersonal_os_get_notepersonal_os_create_notepersonal_os_update_notepersonal_os_complete_notepersonal_os_uncomplete_notepersonal_os_archive_notepersonal_os_restore_notepersonal_os_delete_note
Unified Items
personal_os_list_inbox_itemspersonal_os_list_archive_itemspersonal_os_list_planning_items
Reviews
personal_os_list_reviewspersonal_os_list_review_typespersonal_os_get_reviewpersonal_os_start_reviewpersonal_os_update_review_sectionpersonal_os_apply_review_actionpersonal_os_skip_review_itempersonal_os_complete_reviewpersonal_os_abandon_reviewpersonal_os_get_review_summary
Safety metadata
MCP annotations use only SDK-supported fields: readOnlyHint, destructiveHint, idempotentHint, and openWorldHint.
Confirmation is recommended for:
- every permanent delete tool;
- bulk Task update;
- moving a Task to Trash;
- archive actions;
- abandoning a Review;
- Review item actions because an action may mutate or delete its source.
Lifecycle meanings:
- Trash: recoverable Task soft delete;
personal_os_restore_taskcan recover it. - Archive: reversible removal from active views; the matching restore/unarchive action reactivates it.
- Final delete: permanent and irreversible; applies to Containers, Projects, Collections, Tags, and Notes.
The MCP protocol annotations inform compatible hosts, but the server does not create a custom Accept/Reject UI.
HTTP and errors
The centralized client sends Accept: application/json, JSON bodies where applicable, and a redacted Bearer token header. It safely joins base paths, omits null/empty query values, applies a configurable timeout, validates JSON response shape, and returns structured tool errors:
authenticationfor 401authorizationfor 403not_foundfor 404 or foreign-owned IDsvalidationwith field errors for 422rate_limitedfor 429upstreamfor 5xxnetwork,timeout,invalid_json, orcontractfor transport/response failures
Only transient connection/timeout failures on GET requests are retried, at most twice after the initial attempt. Mutations are never retried automatically.
Security boundary
- One local token represents exactly one Personal OS User.
- The token must have the
ai:accessability andai-api:name prefix. - Normal application tokens cannot call the AI namespace.
- AI tokens cannot call normal User/admin endpoints.
- Authorization values are never included in tool errors.
.envand.env.*are ignored except.env.example.
Known limitations
- Local
stdioonly; no Streamable HTTP transport or remote deployment. - No OAuth or multi-user token selector.
- No Reminders, Recurrences, or Checklist mutation tools.
- No ChatGPT widget, custom frontend, or Accept/Reject UI.
- Tool schemas are maintained as a typed projection of Laravel routes, Form Requests, enums, Resources, and capabilities because Personal OS currently publishes no OpenAPI document.
Future stages may add the explicitly deferred domains, HTTPS remote deployment, and ChatGPT integration after their security and deployment decisions are approved.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。