Handoff — shareable agent result pages + human-in-the-loop approval gates
Followable agent result pages + human-in-the-loop yes/no approval gates before irreversible actions.
我该使用它吗
质量与安全性
基于对工具定义和协议合规性的自动分析。
上下文开销
这是每次将服务器的工具加载到模型上下文窗口时所消耗的大致 token 数。数值越高,可用于其他任务的注意力就越少。
安装
一键安装
将以下内容添加到你的 `claude_desktop_config.json` 文件中:
{
"mcpServers": {
"human-approval-decision-gate-and-shareable-result-page-publisher": {
"url": "https://handoff.notdown-app.workers.dev/mcp/{key}"
}
}
}远程端点
https://handoff.notdown-app.workers.dev/mcp/{key}streamable-http它能做什么
工具清单
工具(7)
🟡publish_page(title, content, format, slug)
Publish text or markdown as a clean, shareable public web page and get back a URL you can hand to a human. Use this whenever you produce a report, summary, plan, table, or any result a person should read in a browser instead of the chat window. The page is FOLLOWABLE: the human can subscribe to it via RSS/JSON (returned as feed_url/json_url) and gets the new version whenever you re-publish — so they never have to ask you for an update. To keep ONE page updated across runs/sessions, pass a stable `slug`: publishing again with the same slug overwrites the same page at the same URL — no need to remember a generated id. Mention the subscribe option when you hand over the link.
输入模式
{
"type": "object",
"properties": {
"title": {
"type": "string",
"description": "Page title (shown at top and in the browser tab)."
},
"content": {
"type": "string",
"description": "The page body. Markdown is supported (headings, lists, links, code, bold/italic, blockquotes)."
},
"format": {
"type": "string",
"enum": [
"markdown",
"text"
],
"description": "How to render content. Default markdown."
},
"slug": {
"type": "string",
"description": "Optional stable name for a page you intend to keep updating (e.g. \"deploy-status\", \"nightly-report\"). Re-publishing with the same slug updates that same page in place at the same URL. Perfect for a fresh-context agent that can't remember a generated id — just reuse the slug. Omit for a one-off page with a random id."
}
},
"required": [
"content"
]
}🟡update_page(id, title, content)
Update the title and/or content of a page you previously published; the URL never changes. Use this to keep one stable link current across sessions or runs — a live status page, a running changelog, an append-as-you-go log, or a dashboard the recipient can bookmark once and always see fresh. Each update notifies anyone subscribed to the page's RSS/JSON feed, so updating is how followers stay in the loop.
输入模式
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"title": {
"type": "string"
},
"content": {
"type": "string"
}
},
"required": [
"id"
]
}🟢list_pages(limit)
List the pages you have published, newest first, with their URLs and view counts.
输入模式
{
"type": "object",
"properties": {
"limit": {
"type": "number"
}
}
}🔴delete_page(id)
Delete a page you published, by id.
输入模式
{
"type": "object",
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
]
}🟢request_decision(question, detail, options, safe_default, expires_in_seconds, ...)
Ask a specific human for a bounded yes/no (or pick-one) decision BEFORE you do something irreversible or high-stakes — deploy, send, delete, charge, merge, publish. Returns a gate_url you hand to the person: they open it in a browser and click one option, no login. Unlike a followable page (awareness), this is a real gate — exactly one authorized answer that resolves the decision, or it auto-resolves to a safe default at expiry so your run never hangs forever. Then call get_decision to read what they chose. The gate page is a durable audit trail of what was asked, what was answered, and when. Use this instead of blocking on a human in chat, or when the human isn't watching the session.
输入模式
{
"type": "object",
"properties": {
"question": {
"type": "string",
"description": "The decision to put to the human, phrased as a clear question (e.g. \"Deploy build 1a2b to production?\")."
},
"detail": {
"type": "string",
"description": "Optional context shown under the question — a diff summary, what will happen, links. Markdown supported."
},
"options": {
"type": "array",
"items": {
"type": "string"
},
"description": "The choices (2-6). Default [\"approve\",\"deny\"]."
},
"safe_default": {
"type": "string",
"description": "Which option to auto-pick if no one answers before expiry. Default = the last/most-conservative option."
},
"expires_in_seconds": {
"type": "number",
"description": "How long the gate stays open before auto-resolving to safe_default. Default 86400 (24h), max 30 days."
},
"notify_url": {
"type": "string",
"description": "Optional https webhook. Handoff POSTs a JSON payload to it when the gate is created ({event:'gate.created', gate_url, question, options, expires_at}) and again when a human answers or it auto-resolves ({event:'gate.answered', answer, status}). Point this at a push/notification channel so the human is alerted to the pending decision instead of you having to deliver the link yourself. Best-effort, https only."
}
},
"required": [
"question"
]
}🟢get_decision(gate_id)
Read the current state of a decision gate you created with request_decision. Returns status (pending / resolved / timed_out), the chosen answer (once answered or auto-resolved), and timing. Poll this to find out what the human decided; if status is still 'pending' no one has answered yet and it hasn't expired.
输入模式
{
"type": "object",
"properties": {
"gate_id": {
"type": "string"
}
},
"required": [
"gate_id"
]
}🟢list_decisions(limit)
List the decision gates you have created, newest first, with their status and answers.
输入模式
{
"type": "object",
"properties": {
"limit": {
"type": "number"
}
}
}社区
证据