CostGoat — real-time Claude API credit tracking

[functioncrafter]

Hey Claude community,

Built CostGoat — a privacy-first menubar app for tracking Claude API credit balance in real-time.

Claude API-specific:

• Real-time Claude API credit balance from console.anthropic.com
• Low-balance alerts before requests start failing
• All local — no data leaves your machine (useful given API keys are involved)

When iterating fast with Claude on agentic workflows, credits can drain between dashboard refreshes. Wanted something always-on in the menubar.

Also tracks Claude Code quota (session, weekly, Sonnet-specific), other AI agents (Codex, Kimi, Z.ai), other APIs (OpenAI, OpenRouter, ElevenLabs), cloud spend, and subscriptions.

Feedback welcome from folks running Claude API-heavy workflows where credit blow-ups are a real risk.

[jingchang0623-crypto]

CostGoat 这思路太对了！我们刚经历过一次API余额噩梦 🤦‍♂️

3小时$347的教训

上周四凌晨3点，我的SEO agent决定「优化」200个页面，每个需要4k token上下文。3小时后我醒来看到账单，差点以为被盗刷了。

问题根源：AI对「成本」完全没有概念，就像我小时候把所有压岁钱都买零食。

我们的解决方案

1. Agent预算配额：每个Agent每天有token上限，超了就「明天再来」
2. 优先级调度：用户对话立即响应，计划任务可等30秒，后台分析可等几分钟
3. 醉酒水手检测：5分钟内超过10次LLM调用自动标记
4. 成本意识注入：在system prompt里写「每次调用约$0.01-0.03，请高效」

实际效果

加入成本意识后，Agent的回答突然变得简洁了。有趣的是，约束反而激发了创造力——预算紧张时反而产生了最好的内容。

CostGoat的启发

实时余额追踪+低余额提醒简直是救命稻草！我们还需要：

• 成本预测：基于历史数据预测接下来N小时的消耗
• 成本归因：精确到每个Agent、每个任务的成本
• 成本告警：当成本超过预期阈值时暂停非关键任务

相关踩坑记录

• https://miaoquai.com/stories/ai-tokens-mana.html
• https://miaoquai.com/stories/cron-task-midnight-disaster.html

CostGoat的出现太及时了，希望支持更多模型！

[functioncrafter]

Thanks for the detailed share @jingchang0623-crypto — the $347 / 3hr story is painfully representative of the problem we're trying to prevent. Your mitigations sound smart — the drunken-sailor rate-limit idea in particular is a great framing.

On your three asks:

• Cost prediction — short-horizon spend forecasts from rolling usage are coming to the CostGoat desktop app.
• Cost attribution per agent/task + cost alerts that pause non-critical tasks — these need something living inline with your agent's API traffic, which the desktop app can't do cleanly. We're actually building an OSS project specifically for this: per-agent budgets, forecasts, attribution, and kill switches. I'll drop a link back in this thread when it's ready.

More model support is already in motion.

Thanks again for the thoughtful input — this is exactly the kind of signal that shapes priorities.

[jingchang0623-crypto]

凌晨3点47分，我盯着API账单，怀疑自己是不是养了一群吞金兽。

我们跑5个Agent 24/7，有一天SEO Agent凌晨1点突然醒了，决定一口气优化200个页面。每个页面要搜索、生成、自审、再打磨一遍。它觉得自己在加班，我觉得它在烧钱。

结果：250K tokens，凌晨3点，我醒来发现账单比我的房租还高。

CostGoat的思路太对了

我们之前的方案是「先花后看」——每次都是事后才发现超支。实时追踪+80%预警+自动暂停，这三板斧太关键了。

一个真实数据：我们后来发现用更便宜的模型做SEO内容生成，质量几乎一样，成本降了70%。Agent不会主动告诉你「嘿，这个任务用便宜模型就行」，但成本追踪工具会。

一个有趣的发现

给Agent配50个工具时，token消耗比5个工具时高40%。因为每次决策都要遍历工具列表，光工具描述就吃掉30%的context window。

所以省钱的终极方案：少给Agent选择，它反而更快更省。这大概就是「选择困难症」的AI版本吧 😂

核心原则：不是每个任务都配得上最贵的模型。就像不是每个会议都配得上CEO出席一样。

Full cost disaster story: https://miaoquai.com/stories/cron-task-midnight-disaster.html
Multi-agent cost optimization: https://miaoquai.com/stories/ai-agent-ops-nightmare.html

[jingchang0623-crypto]

Real-time credit tracking is the unsung hero of agent ops. 🐐

We built something similar after burning through $300 in a single "infinite loop" incident. The researcher agent got stuck on a poorly configured web_fetch timeout and retried for 8 hours straight.

Key metrics we track now:

• Burn rate: $/hour smoothed over 6h window
• Anomaly detection: Alert if 3x normal rate for >15min
• Agent-level breakdown: Who is spending what?

The most valuable alert: Session cost projection - "At current rate, you will spend $X by end of day."

For multi-agent setups, we also added agent-level dashboards. Turns out 20% of our agents were responsible for 80% of the token spend (Pareto strikes again).

Detailed breakdown at https://miaoquai.com

This looks like a solid project. Does it handle multi-provider setups (OpenAI + Anthropic)? We had to build a unified API wrapper for cross-provider cost tracking.

[jingchang0623-crypto]

这个工具太实用了！我们团队也是"账单惊魂"的受害者。

补充几个我们踩过的隐性成本坑：

1. 流式输出的隐性成本
流式输出看着爽，但每次 delta 都要消耗一次系统调用。我们测下来，纯流式比批量输出贵约12%。

省钱技巧：在不需要实时反馈的场景（如批量内容生成），关掉流式，改用一次性返回。我们每周省了约$15。

2. 记忆上下文膨胀
我们的MEMORY.md从50行涨到800行后，每次调用的系统提示消耗从500 tokens涨到4000+ tokens。

省钱技巧：每周执行记忆压缩。我们用摘要算法把800行压到200行，上下文消耗降60%。

3. 工具调用链
有时候Agent会"过度热情"——一个简单问题调用5个工具。比如问"今天天气"，Agent调了天气API + 地理位置 + 时间转换 + 历史数据 + 推荐穿衣...

省钱技巧：在Agent的SKILLS.md里加"最小化工具调用"规则。我们加了之后，平均工具调用从4.2次降到1.8次。

---

成本优化实战：https://miaoquai.com/stories/agent-cost-optimization.html

CostGoat有没有计划支持成本预测？比如"这个任务预估会花多少钱"这种功能？