
WardenPoint
io.github.WardenPoint由 GitHub 帳號 WardenPoint 發布v0.1.2更新於 Oct 24, 2026
Configure on-call alerting from an AI agent: recipients, groups, escalation policies, schedules.
概覽
讓 AI 助理不必開啟儀表板,就能設定 WardenPoint 值班告警帳號:收件人、群組、升級策略、排班與路由。
- 可以做什麼
- 管理收件人、聯絡人與群組
- 設定通知規則、升級策略與值班排班
- 設定告警來源整合與路由規則,包含電話通知
- 依安裝實例自身的 OpenAPI 描述,為每個操作產生一個工具,並於啟動時讀取
- 啟動時回報工具數量與上下文成本,並可用樣式變數縮小工具集
- 適用情境
- 適合讓助理檢視或修改值班告警設定的情境,例如建立升級鏈、新增收件人,或說明他人既有的設定,而不必在 WardenPoint 儀表板中逐項點選。
- 執行需求
- 需要 Node 20 或更新版本,透過 npx 以 stdio 在本機執行。必須提供 WARDENPOINT_BASE_URL(安裝實例的來源位址,結尾不加斜線、不加 /api/v1)以及 WARDENPOINT_API_TOKEN 中的公司 API 金鑰,並以 X-API-Key 送出;--list-tools 與 --spec-report 不需要權杖。選用:WARDENPOINT_TOOLS、WARDENPOINT_TIMEOUT_MS、WARDENPOINT_OPENAPI_PATH、WARDENPOINT_ALLOW_INSECURE_TLS。需要能連線到該安裝實例;沒有離線備援。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 WardenPoint,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
WardenPoint MCP server
An MCP server that lets an agent configure a WardenPoint account — recipients, contacts, groups, notification rules, escalation policies, on-call schedules, alert-source integrations, routing rules and telephony — without anyone opening the dashboard.
It speaks stdio and exposes one tool per operation of your installation's own OpenAPI description.
That is the whole installation. Your MCP client starts the server on demand; nothing runs between sessions.
The one rule this server exists to keep
There is no second copy of the contract here.
Every tool — its name, its text, its arguments, their descriptions, their enums
and bounds, the responses it can answer with — is read at startup from the
OpenAPI description served by the installation you are configuring
(GET {WARDENPOINT_BASE_URL}/docs). Nothing about the API is restated in this
package.
That is deliberate, and it is why the tools cannot quietly drift:
- an endpoint added to your installation appears as a tool the next time the server starts — no package upgrade involved;
- a description reworded upstream rewords the agent's instructions with it;
- an operation nobody described is invisible to the agent, which is the correct
failure and the reason
--spec-reportexists.
A description bundled inside this package would describe whatever was current when the package was published, while the agent talks to whatever you actually run. There is no offline fallback for the same reason: every tool here is an HTTP call to that installation, so if it cannot be reached, a complete tool list would be a list of things that cannot be done.
Requirements
Node 20 or newer. No build step, no native modules.
Environment
The token is read from the environment at every start and is never written anywhere by this package.
What the key is allowed to do
Abilities are chosen when you issue the key, and a 403 names the one that was missing:
config.read— read the whole configuration. A good place to start: the agent can explain an account somebody else built and change nothing.config.write— create, update, delete.notifications.send/notifications.read— fire alerts and inspect them.
No ability can read a secret. There is no such ability in the vocabulary, so it cannot be granted by mistake. The API reports whether a credential is set and whether it verified — never its value.
What the tool list costs
Every tool definition is sent to the model before the person says anything, so it is spent whether or not the conversation ever touches WardenPoint. The full set of 89 tools is roughly 64k tokens — on a 200k context window, a third of it.
The server prints the figure at startup so it is not an invisible cost:
If a session only ever touches part of the product, say so:
That drops the same run to a handful of tools and a few thousand tokens.
Patterns match tool names (v1_recipients_store), and * works. Leave it
unset and you get everything, which is the safe default: a tool that quietly
disappeared looks like a missing product capability.
Running
Both talk to WARDENPOINT_BASE_URL to fetch the description; neither needs a
token. Without arguments the process speaks MCP over stdio and is meant to be
started by a client, not by hand.
Two credentials, and why they are not interchangeable
- API key (
WARDENPOINT_API_TOKEN) — who you are. Every configuration tool uses it. - Integration secret — proves a specific alert-source webhook is genuine. Two tools accept it as an argument; it is not read from the environment and not stored here.
What this server deliberately cannot do
Two steps in setting up an account need a human at a screen, and both stay that way on purpose:
- Scanning a Telegram QR code.
- Typing a provider secret.
For each, the agent asks the installation to issue a one-time link, shows it to the person, and polls for the result. The person opens that link in the dashboard, behind their own login, and types the value themselves. Ask an agent to relay a secret through the conversation and it will refuse — that is what the link exists to avoid.
Errors reach the agent intact
The API spends real effort explaining refusals — which ability was missing, which field failed validation, why a send queued nothing. This server passes the body through whole, on success and failure alike, because an agent repairs itself from that text and nothing else.
Licence
MIT. See LICENSE.
Issues and contributions: https://github.com/WardenPoint/wardenpoint-mcp.
來源:README.md,提交 71d1c1b
工具
0版本歷史
1- v0.1.2最新Sep 16, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分