Agent Dash
A self-hosted push inbox for AI agents that enables agents to send updates and ask questions, waiting for user answers before continuing.
README
Agent Dash
Your agents report here. A self-hosted push inbox for AI agents — deploy your own on Cloudflare's free plan, and any agent (Claude Code, Codex, Cursor, Antigravity, Kimi, or a raw ChatGPT/Claude chat) can send updates to your phone and ask you a question and wait for your answer before continuing.
Think "ntfy for agents, with a reply button." Vendor-neutral, single-user, free to run.
<p align="center"><img src="public/icon.svg" width="96" alt="Agent Dash"></p>
Why
Every AI vendor has a companion app — but it only talks to their agent. When you kick off a long research run, a refactor, or a build in Claude Code / Codex / Cursor, you either babysit the terminal or miss the moment it finishes or gets stuck.
Agent Dash is one inbox for all of them:
- 🔔 Push updates — milestones and completions land as notifications on your phone (installable PWA, works on iOS + Android).
- ❓ Ask-and-wait — an agent can post a question (a choice, or a whole form) and poll until you answer, then continue. "I finished the research — VC or customer framing for the deck?" → you tap → it keeps going.
- 🧱 Structured messages — agents send typed blocks (markdown, progress bars, tables, forms), not raw text. Rendered safely, never as HTML.
- 🔌 Connect anything — a portable skill, a one-line MCP server, an OpenAPI spec for ChatGPT Actions, and drop-in Claude Code hooks.
- 🆓 Free to host — one Cloudflare Worker + D1 + KV. No Durable Objects, no Postgres, no Firebase. Fits comfortably in the free tier for one person.
Deploy your own (≈5 minutes)
You need a free Cloudflare account and Node 20+.
git clone https://github.com/Prajeevan/agent-dash.git
cd agent-dash
pnpm install # or npm install
npx wrangler login # authorize wrangler with your Cloudflare account
# Create the database + KV, then wire their ids into wrangler.jsonc:
npx wrangler d1 create agent-dash
npx wrangler kv namespace create SESSIONS
# → paste the printed database_id and id into wrangler.jsonc
pnpm setup # generates keys, sets secrets, migrates DB, deploys
pnpm setup prints your worker URL, a magic login link, and your
agent connection snippet. Then:
- Scan the QR code it prints with your phone (or open the link) → Add to Home Screen → open the app → Settings → Enable notifications.
- Give an agent the MCP snippet or the skill.
Re-run pnpm run login any time for a fresh 15-minute login link + QR (rendered
locally in your terminal — the magic-link token never leaves your machine). Re-run
pnpm setup --rotate to regenerate all keys.
Connect an agent
MCP (Claude Code, Cursor, Codex, any MCP client) — one entry:
{
"mcpServers": {
"agent-dash": {
"url": "https://agent-dash.your-name.workers.dev/mcp",
"headers": { "Authorization": "Bearer YOUR_AGENT_KEY" }
}
}
}
Tools: notify(title, blocks?, priority?), ask(title, blocks), wait_for_answer(question_id).
Skill — install it into any Agent-Skills runtime (Claude Code, Cursor, Codex, …) straight from GitHub:
npx skills add Prajeevan/agent-dash
This drops skills/agent-dash/SKILL.md into your agent so it knows how to reach
your hub; then give it your URL + AGENT_KEY.
curl (anything that can make an HTTP request, including a raw chat) — the one-liner:
curl -X POST "$AGENT_DASH_URL/api/v1/events" \
-H "Authorization: Bearer $AGENT_KEY" -H "Content-Type: application/json" \
-d '{"agent":"claude","title":"Build finished","priority":1}'
CLI — npx agentdash connects an agent, logs you in on your phone
(QR), and sends updates/questions from scripts:
npx agentdash login # save + verify your hub URL and agent key
npx agentdash connect # write ./.mcp.json for your agent
npx agentdash notify "Build finished" --priority 1 --project "API"
ChatGPT Actions — import https://your-url/api/v1/openapi.json.
Claude Code hooks — drop-in start/finish/needs-input pushes.
How it works
Agents ──POST /api/v1/events─────▶ Cloudflare Worker (TanStack Start SSR + API)
──POST /api/v1/questions──▶ ├─ /api/v1/* agent REST (bearer AGENT_KEY)
◀─GET /api/v1/questions/:id ├─ /mcp stateless MCP
├─ /app dashboard PWA (magic-link session)
You (PWA) ◀── Web Push (VAPID) ─── └─ Bindings: D1 (events/questions/subs), KV (sessions)
──poll feed while open─▶ Cron: expire questions + prune old events
- No Durable Objects. Delivery is Web Push (real-time, free) plus polling — agents check for answers every ~10s; the dashboard polls a cursor feed only while a tab is open. A single-user workload stays well inside free limits.
- Two credentials, separate concerns.
AGENT_KEY(agents post/poll) andAPP_SECRET(signs your magic-login links + sessions). They rotate independently. Sessions live in KV with a TTL; "log out everywhere" bumps an epoch that invalidates them all. - Blocks, not HTML. Agent messages are
zod-validated typed blocks; the UI renders a known set — no agent-supplied markup is ever executed.
See PLAN.md for the full design rationale.
Tech
TanStack Start (React 19) on Cloudflare Workers · D1 · KV · Web Push (hand-rolled VAPID + RFC 8291 on WebCrypto, no Node deps) · Vite · TypeScript.
Dev: pnpm dev. Typecheck: pnpm typecheck. Deploy: pnpm run deploy.
Instant mode (optional)
Delivery is polling by default (free tier, no Durable Objects). To get
sub-second live updates instead, flip one flag: set "INSTANT": "1" in
wrangler.jsonc and pnpm run deploy. That activates a single hibernating
Durable Object — hibernated sockets don't bill duration, so it stays free-tier
friendly for one user. The dashboard automatically switches from polling to a
live WebSocket. Set it back to "0" to disable.
Roadmap
- [x] Instant mode via a hibernating Durable Object (opt-in flag above).
- [ ] Capacitor shell for native app-store builds where PWA push is inconvenient.
- [ ] More block types (charts, diffs) — PRs welcome (raw-HTML blocks will be declined by design).
License
MIT — see LICENSE. The single-user core stays MIT forever.
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。