AskHumanToWork MCP Server
Enables AI agents to capture, manage, and retrieve todos with due dates and provenance, while automatically escalating reminders until tasks are completed.
README
AskHumanToWork
The todo hub where your AI asks you to work.
AI agents (Claude Desktop, Claude Code, any MCP client) capture todos into your list — with due dates and provenance ("why this exists") — and AskHumanToWork reliably reminds you until they're done. Optionally (Pro), todos mirror out to Microsoft To Do / Google Tasks so they appear where you already look.

Why
Heavy AI users generate implicit todos constantly: "you should follow up on X", "remember to deploy Y on Friday". Those evaporate in chat history. AskHumanToWork captures them at the source — the AI conversation itself — and closes the loop with an escalating reminder engine.
Every AI-created todo shows who added it and why:


Tutorial — zero to your first AI-captured todo
0. Prerequisites
- Node ≥ 20 and pnpm (
npm i -g pnpm) - PostgreSQL 16, Redis, and Mailpit (local email catcher)
- Flutter 3.x (only for the mobile app)
macOS one-liner for the services:
brew install postgresql@16 redis mailpit
brew services start postgresql@16 redis mailpit
createdb askhumantowork
(Prefer containers? docker compose up -d starts the same three services.)
1. Install & configure
git clone https://github.com/KuronokiCorp/AskHumanToWork.git && cd AskHumanToWork
pnpm install
cp .env.example .env
# In .env, set:
# ENCRYPTION_KEY → run: openssl rand -base64 32
# VAPID keys → run: npx web-push generate-vapid-keys (for browser push; optional)
2. Build, migrate, seed
pnpm build
pnpm db:migrate
pnpm db:seed
The seed prints your demo credentials — copy the MCP token now, it's stored only as a hash:
Seeded demo data.
Login: demo@askhumantowork.local / demo1234
MCP token: tfa_XXXXXXXXXXXXXXXXXXXXXXXX
3. Run it (3 terminals)
pnpm dev:api # API on :3000, MCP endpoint at /mcp
pnpm --filter @askhumantowork/api dev:worker # reminder + sync engine
pnpm dev:web # web app on :5173
Open http://localhost:5173 and sign in with the demo account. You'll see seeded todos, including AI-captured ones in the AI Inbox.
4. Add your first todo (human style)
In the quick-add box, natural language just works:
Ship the release notes @friday 5pm #Work !2
@… = due date (resolved in your timezone, server-side) · #… = project (auto-created) ·
!1-3 = priority.
5. Connect Claude — the main event
claude mcp add askhumantowork \
--env TODO_API_TOKEN=tfa_your_token_here \
--env TODO_API_URL=http://localhost:3000 \
-- node <repo>/packages/mcp/dist/bin.js
(Once the askhumantowork-mcp package is published to npm, replace the last line with
-- npx -y askhumantowork-mcp. Claude Desktop users: see the equivalent JSON config in
packages/mcp/README.md.)
Now, in any Claude session:
You: remind me to review the auth PR tomorrow at 3pm
Claude: (calls
add_todowithdue_natural: "tomorrow 3pm"andorigin_context: "You asked while we were discussing the auth refactor.") Added — it'll remind you tomorrow at 2pm and 3pm. → http://localhost:5173/t/…
Refresh the AI Inbox — the todo is there, with Claude's name and the reason it exists.
Also try:
- "what's on my plate today?" → Claude calls
get_agenda - "I finished the auth PR" →
complete_todo(reminders cancelled) - The
/capture-followupsprompt → Claude scans the whole conversation and files every commitment as a todo.
6. Watch a reminder fire
Reminders ladder automatically: 1 day before → 1 hour before → at due → daily overdue nudges. To see one in 20 seconds:
curl -X POST http://localhost:3000/api/todos \
-H "Authorization: Bearer tfa_your_token" -H "Content-Type: application/json" \
-d '{"title":"Reminder demo","dueNatural":"in 10 minutes","reminders":["in 20 seconds"]}'
Open the Mailpit UI at http://localhost:8025 — the reminder email arrives, including the AI provenance if an agent created the todo. (Enable browser push in Settings → Notifications for native notifications; quiet hours are respected.)
7. Mobile app (Flutter)
cd mobile && flutter run # iOS simulator or Android emulator
Sign in with the demo account. Same Today/Upcoming/Projects/AI Inbox views; server reminders are mirrored as local notifications; completing a todo on the phone syncs back instantly.
8. Mirror to Microsoft To Do / Google Tasks (Pro)
Third-party sync is a Pro-plan feature. As the seeded admin you're already Pro; upgrade other users on Settings → Admin.

- Register an OAuth app —
Microsoft: Azure Portal → App registrations, delegated
Tasks.ReadWrite+offline_access, redirect URIhttp://localhost:3000/api/integrations/ms-todo/callback· Google: Cloud Console → OAuth credentials, scopehttps://www.googleapis.com/auth/tasks, redirect URIhttp://localhost:3000/api/integrations/google-tasks/callback - Paste client id/secret in Settings → Admin (or the
.envvars). - Settings → Integrations → Connect, pick a target list, direction (two-way / outbound), and filters (AI-only, priority threshold).
AskHumanToWork stays the source of truth: mirrors degrade per provider capability (Google Tasks has date-only due dates and no reminders — our reminder engine still covers you), and completing a task in the external app completes it here and cancels the reminders (2-minute polling).
Reference
Architecture
AI agents ──MCP (stdio / HTTP)──►┐
Web (React) ──REST──────────────►│ Fastify API ──► PostgreSQL (source of truth)
Flutter app ──REST──────────────►┘ │
├─► BullMQ/Redis: reminder ladder + sync outbox
├─► Email (SMTP) + Web Push reminders
└─► Adapters: Microsoft To Do, Google Tasks (Pro)
| Package | What |
|---|---|
packages/shared |
zod schemas, enums, server-side natural-language date resolution |
packages/db |
Drizzle ORM schema, migrations, seed |
packages/core |
domain services, entitlements, sync engine, provider adapters |
packages/api |
Fastify REST API, auth, /mcp HTTP transport, background workers |
packages/mcp |
askhumantowork-mcp — publishable stdio MCP connector |
packages/web |
React web app (Vite + Tailwind + TanStack Query) |
mobile/ |
Flutter app (Riverpod + dio + local notifications) |
MCP surface
Tools: add_todo (natural dates, origin_context provenance, idempotent, sync_to routing) ·
list_todos · search_todos · update_todo · complete_todo · reschedule_todo · get_agenda ·
list_projects · list_integrations · resolve_time
Resources: todo://agenda/today · todo://agenda/overdue · todo://projects
Prompts: capture-followups · review-my-todos
Remote clients can skip the local install entirely: Streamable HTTP MCP at POST <server>/mcp
with Authorization: Bearer <token>.
Auth model
Web = cookie sessions. Mobile = long-lived device tokens (POST /api/auth/login with
mode:"token"). AI agents = scoped Personal Access Tokens (todos:read/write, projects:read,
integrations:read) created in Settings → API tokens.
Plans
Everything is free except third-party sync (Pro). Gating is enforced server-side (connect → HTTP 402, no outbound fan-out, inbound pollers skip). Until billing ships, admins set plans on the Admin page.
Deploying
Single-image deploy — the API container also serves the built web app (SERVE_WEB=true) and runs
migrations on start; a second container from the same image runs the reminder/sync worker:
cp .env.example .env # set SESSION_SECRET + ENCRYPTION_KEY (+ VAPID keys)
docker compose --profile app up -d --build
# → http://localhost:3000 (web + API + MCP), worker running, Postgres/Redis/Mailpit included
For a real host (Fly.io / Railway / a VPS): build the Dockerfile, provide DATABASE_URL,
REDIS_URL, SMTP credentials, and set COOKIE_SECURE=true + TRUST_PROXY=true behind HTTPS.
Sessions and rate limits are Redis-backed, so multiple API instances are safe.
Publishing the npm connector
cd packages/mcp && npm login && pnpm publish
Ships only dist/ with two runtime deps; prepublishOnly builds automatically.
Tests
pnpm typecheck && pnpm test # TS packages (incl. date-resolution unit tests)
cd mobile && flutter analyze # Flutter
A scripted end-to-end regression (auth, todos, dedup, agenda, MCP both transports, reminder
delivery/cancellation, plan gating) lives in the feature-tester agent charter at
.claude/agents/feature-tester.md — Claude Code users can run it with "run the feature-tester".
Roadmap
Recurring todos · edit-in-place UI · web search & tag filters · notification action buttons · FCM/APNs mobile push · morning AI digest · Graph webhooks (realtime inbound) · Todoist adapter · Stripe billing · cloud deploy.
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。