
HyperProbe 是由 Y Combinator 投資的 AI 原生生產環境除錯工具。後端團隊可在不重新部署、不重啟服務的前提下,於正在運行的線上程式碼放置唯讀虛擬中斷點,從告警追到有證據支持的根因分析。
HyperProbe 是什麼?白話說明
簡單講,HyperProbe 解決的是:線上出問題時,日誌裡往往沒有那個「關鍵變數」的值。
傳統流程是:猜 → 加 log → 等 CI/CD → 再部署 → 等故障重現。HyperProbe 改在懷疑出錯的那一行掛「探針(probe)」:下一筆真實流量經過時拍下變數快照,然後自動消失。服務不暫停、使用者無感、也不用為了除錯改程式發版。
可視為給生產環境用的非阻塞除錯器。Cursor、Claude Code 等程式智能體透過 MCP 遠端下探針、讀快照、提修復建議。
它解決哪些痛點
靜默失敗、堆疊誤導、例外被吞、競態、第三方 API 變更、AI 寫碼後的線上盲區等。
虛擬探針與安全
唯讀、非阻塞、TTL/速率/命中上限、程序內 PII 脫敏、不可篡改稽核日誌。
執行環境:JavaScript、TypeScript、Java、Python、Ruby。
AI 值班流程
對接 PagerDuty、Datadog、Slack → 規劃 → 下探針 → 捕獲 → 確認 RCA。
定價(2026 年 9 月核實)
按服務計費,探針與捕獲不限量。
- Free: $0,1 服務,託管雲,7 天歷史,含 MCP
- Professional: $99/服務/月(年付 $79,最少 3 服務)
- Enterprise: 客製年約,自建/VPC,審批閘門
詳見定價頁。
常見問題
除錯生產環境一定要重新部署嗎?
不用。SDK 在行程內動態注入探針。
探針有次數限制嗎?
沒有按量計費,看接入的服務數。
能全部跑在自有基礎設施嗎?
Enterprise 可以,控制面與捕獲儲存在你的 VPC。
官方資源
Installation and first capture
Paste one prompt into your coding agent with the installation guide. The agent installs the in-process SDK, wires your entrypoint, enables source maps, and registers the MCP server. Official docs target 15 minutes to first capture. Architecture: SDK agent inside your service plus a broker/control plane (managed or self-hosted).
Integrations: PagerDuty, Datadog, Slack for alerts; Cursor, Claude Code, Codex, Opencode via MCP for agent-driven debugging; VS Code extension for manual probes.
Comparison: with vs without HyperProbe
Without HyperProbe, engineers grep logs, add print statements, wait for CI/CD, redeploy, and hope the failure reproduces—often 2+ hours for a one-line fix. With HyperProbe, the agent locates the frame, places a probe in background, captures on next matching request, and confirms RCA in under 10 minutes in published examples—without redeploy cycles.
More FAQ
What counts as one service?
One application with the SDK installed. Forty Kubernetes pods still count as one service. HyperProbe does not bill per host, container, traffic volume, or seat.
Can we start managed and move to self-hosted?
Yes. The SDK is identical; self-hosted changes where the control plane runs, not your instrumentation.
Which alert sources are supported?
PagerDuty, Datadog, and Slack per the product website. The agent reads logs and distributed traces to plan probe placement.
Does HyperProbe replace observability tools?
No. It complements logs and traces by capturing variable state those signals miss—especially silent failures and logic bugs without exceptions.






