dum-e
Enables AI agents to manage tasks through a structured lifecycle (todo, in_progress, ai_testing, manual_testing, deployment, completed) with human approval gates, providing a single-user task management system with CLI, REST API, and live updates.
README
<p align="center"> <img src="assets/logo.svg" alt="dum-e" width="96" /> </p>
dum-e
Named for Tony Stark's clumsy robot arm. A personal, single-user AI task manager that AI agents (Claude, Codex, opencode, any LLM) drive through a defined lifecycle, with you as the approval gate.
Three surfaces, one shared core: a powerful CLI, a REST API with live updates, and an MCP server that any agent can plug into.
The idea
Whenever you delegate work to an AI agent, it first creates a task here. The agent pulls the highest-priority task from the queue, analyzes it, works it, runs AI-level testing, then hands it to manual testing. It waits there for your approve or reject. Rejected tasks bounce back with high-priority comments the agent must address; approved tasks flow to deployment (PR review and merge) then completed. The loop repeats until you approve. Tool- and LLM-agnostic.
agent auto-advances human gate
┌──────┐ ┌─────────────┐ ┌────────────┐ ┌────────────────┐ ┌────────────┐ ┌───────────┐
│ todo │ → │ in_progress │ → │ ai_testing │ → │ manual_testing │ → │ deployment │ → │ completed │
└──────┘ └─────────────┘ └────────────┘ └────────────────┘ └────────────┘ └───────────┘
↑ │ │
└────────────── reject ──────────────┴───────────────────┘
(priority bumped, flagged for the agent to fix first)
Every transition requires a comment, recorded in an append-only history.
Install
Requires Node.js 22+ and pnpm 9+.
git clone <repo> dum-e && cd dum-e
pnpm install
pnpm build
Then put dum-e on your PATH. The recommended way is a small launcher that pins the Node the native dependency (better-sqlite3) was built against, so the CLI works no matter which node is first on your PATH (a plain symlink to dist/index.js breaks if your default node differs from the one you built with):
mkdir -p ~/.local/bin
cat > ~/.local/bin/dum-e <<EOF
#!/bin/sh
exec "$(command -v node)" "$PWD/packages/cli/dist/index.js" "\$@"
EOF
chmod +x ~/.local/bin/dum-e
Ensure ~/.local/bin is on your PATH. Alternatively, if you have run pnpm setup (so pnpm has a global bin dir): pnpm --filter @dum-e/cli link --global.
If you rebuild against a different Node version later (e.g. after
pnpm installunder a new node), regenerate the launcher so it points at the matching interpreter.
Verify:
dum-e --version
dum-e --help
Quickstart
dum-e init # create ~/.config/dum-e/{config.yaml,db}
dum-e project add "Web App" -k WEB # a project
dum-e task add "Fix login" --type bug --priority high
dum-e task next -m "starting" # claim the highest-priority queued task
dum-e task move <id> ai_testing -m "unit tests green"
dum-e task move <id> manual_testing -m "verified locally" # now waits for you
dum-e task ls --status manual_testing # what awaits your review
dum-e serve --open # REST API + web UI, opens the browser
Surfaces
- CLI (
dum-e ...): full task and project lifecycle from the terminal. Add--jsonto any command for scriptable output. See docs/CLI.md. - REST + web (
dum-e serve): Hono API with Server-Sent Events for live updates, serving the React web UI (kanban board, table view, metrics, detail drawer). See docs/REST-API.md. - MCP (
dum-e mcp): a Model Context Protocol server on stdio exposing task tools to any agent. See docs/AGENTS.md. Teach an agent to use dum-e by handing it skills/dum-e/SKILL.md.
All three open the same local SQLite database directly (WAL mode: concurrent readers, serialized writes), so none of them needs the others to be running.
Architecture
Modular monolith, pnpm workspaces. Every surface is a thin adapter over @dum-e/core.
| Package | Responsibility |
|---|---|
@dum-e/core |
Domain types, storage (SQLite + FTS5), config, state machine, services, event bus |
@dum-e/server |
Hono REST API + SSE; serves the web build |
@dum-e/mcp |
MCP server (stdio): task tools for any agent |
@dum-e/cli |
commander CLI |
@dum-e/web |
Vite + React: kanban board, table view, charts, detail drawer |
Runtime state lives in ~/.config/dum-e/ (config.yaml, db, attachments/), auto-created on first run. Override the location with $DUM_E_HOME. Full detail in docs/ARCHITECTURE.md.
Configuration
~/.config/dum-e/config.yaml holds the agent roster, model roster, defaults, and server host/port. Show or locate it:
dum-e config show # effective config (YAML, or JSON with --json)
dum-e config path # path to the file
Development
pnpm build # build every package (tsup for libs, Vite for web)
pnpm test # Vitest (state machine + services)
pnpm typecheck # tsc --noEmit across the workspace
pnpm lint # oxlint
After editing the CLI, rebuild it: pnpm --filter @dum-e/cli build. The global dum-e symlink points at the build output, so it picks up changes automatically.
License
MIT
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。