Show HN: WatchLLM – 调试 AI 智能体，分步展示并进行成本归因
1 分•作者: Kaadz•6 个月前
大家好，HN！我开发了WatchLLM来解决我在构建AI智能体时遇到的两个问题：
1. 调试智能体非常痛苦 - 当你的智能体调用20次工具却失败时，很难搞清楚哪个决策是错误的。WatchLLM提供了一个逐步的时间线，显示了每一个决策、工具调用和模型响应，并解释了智能体做出这些行为的原因。
2. 智能体成本飞速上涨 - 智能体喜欢陷入循环或反复调用昂贵的工具。WatchLLM跟踪每一步的成本，并标记异常情况，例如“检测到循环 - 相同操作重复3次，浪费0.012美元”或“高成本步骤 - 0.08美元超过阈值”。
核心功能：
* 包含成本明细的每个智能体决策的时间线视图
* 异常检测（循环、重复工具、高成本步骤）
* 语义缓存，作为额外福利，可节省40-70%的LLM费用
* 适用于OpenAI、Anthropic、Groq - 只需更改你的baseURL
它基于ClickHouse构建，用于实时遥测，并使用向量相似性进行缓存。智能体调试器使用LLM生成的每个步骤发生原因的摘要来解释决策。
目前，每月5万次请求以内免费。我正在寻找早期用户，他们正在构建智能体，并希望更好地了解实际发生的事情（以及它的成本）。
试用：[https://watchllm.dev](https://watchllm.dev)
欢迎提供关于其他有用的调试功能的反馈。当你的智能体行为不当时，你希望拥有什么功能？
查看原文
Hi HN! I built WatchLLM to solve two problems I kept hitting while building AI agents:<p>1. Debugging agents is painful - When your agent makes 20 tool calls and fails, good luck figuring out which decision was wrong. WatchLLM gives you a step-by-step timeline showing every decision, tool call, and model response with explanations for why the agent did what it did.<p>2. Agent costs spiral fast - Agents love getting stuck in loops or calling expensive tools repeatedly. WatchLLM tracks cost per step and flags anomalies like "loop detected - same action repeated 3x, wasted $0.012" or "high cost step - $0.08 exceeds threshold".<p>The core features:<p>Timeline view of every agent decision with cost breakdown
Anomaly detection (loops, repeated tools, high-cost steps)
Semantic caching that cuts 40-70% off your LLM bill as a bonus
Works with OpenAI, Anthropic, Groq - just change your baseURL<p>It's built on ClickHouse for real-time telemetry and uses vector similarity for the caching layer. The agent debugger explains decisions using LLM-generated summaries of why each step happened.
Right now it's free for up to 50K requests/month. I'm looking for early users who are building agents and want better observability into what's actually happening (and what it's costing).
Try it: <a href="https://watchllm.dev" rel="nofollow">https://watchllm.dev</a>
Would love feedback on what other debugging features would be useful. What do you wish you had when your agents misbehave?