正在加载视频...

视频加载失败

Very nice Claude Code dashboard for power users to track every request. - Use multiple subscriptions with load balance to switch accounts - Set models for subagents - Track latency, token usage, and costs in real-time - Full request logging and error traces (Link below!)

110,838 次观看 • 1 年前 •via X (Twitter)

38 条评论

Nifal Adam 的头像
Nifal Adam1 年前

Only for API? A load balance for subscription usage would have been awesome

Ian Nuttall 的头像
Ian Nuttall1 年前

Supports Pro and Max plans too

Nifal Adam 的头像
Nifal Adam1 年前

Damn. Got to check it out. Right now my usage is under $200 but interesting

Raymond Weitekamp 的头像
Raymond Weitekamp1 年前

hey @kieranklaassen this made me think of you. i think @danshipper will approve at least two max subs per @every employee ;)

Ian Nuttall 的头像
Ian Nuttall1 年前

@kieranklaassen @danshipper @every I’ll take an extra max sub too plz

Jess.py 的头像
Jess.py1 年前

this looks cool but youre doing this for learning purpose & the views right? because you will get alot of people pissed , considering we all received the email that says we'll be rate limit soon

Ian Nuttall 的头像
Ian Nuttall1 年前

It’s not mine, I found it on Reddit and liked the project. I think the rate limits will be okay and hopefully will improve how people “vibe code” to be more conservative with tokens!

Kurishita 的头像
Kurishita1 年前

Mind is slightly different. But I can directly start and stop agent tasks and can add new objectives to an agent directly from my dashboard.

Alex 🇵🇹 的头像
Alex 🇵🇹1 年前

Much needed after last announcement from Anthropic 😅

zeo 的头像
zeo1 年前

AI builders: ToS? never heard of it

Karim C 的头像
Karim C1 年前

this is exactly what mario needed when we hit rate limits at 3am the multi-subscription load balancing is chef's kiss. we've been manually swapping API keys like animals real question: does it handle the anthropic vs openai context window differences automatically? mario processes restaurant analytics with varying document sizes and the context juggling gets messy production AI needs this level of observability. most people think AI deployment is just "call API" until their first 429 error at scale

Michael Giannulis 的头像
Michael Giannulis1 年前

Super cool how do you fund this stuff so quick

Ian Nuttall 的头像
Ian Nuttall1 年前

Reddit usually! Or lots of people email me their projects

Michael Giannulis 的头像
Michael Giannulis1 年前

You keep me busy haha

Harun R. 的头像
Harun R.1 年前

Wow this is what power users need

Sunny 的头像
Sunny1 年前

with this you can calculate your usage for that weekly limit?

Ian Nuttall 的头像
Ian Nuttall1 年前

@sunnypause I don’t think so unless Anthropic are transparent with the exact limits

L 的头像
L1 年前

Nice!!!

Kevin Salzmann 的头像
Kevin Salzmann1 年前

Does it also work for pro users?

Ian Nuttall 的头像
Ian Nuttall1 年前

Yup, Pro and Max plans supported

AllAboutAI 的头像
AllAboutAI1 年前

super handy

BEN SYNE 的头像
BEN SYNE1 年前

made by @snipeship by the way 📈

Kevin Schawinski 的头像
Kevin Schawinski1 年前

lol very good

Rob Schmidt 的头像
Rob Schmidt1 年前

Have you tried out Devin AI yet? How does it compare to CC?

Ian Nuttall 的头像
Ian Nuttall1 年前

I haven’t but they emailed me a bunch of times to try it. I will try and test it out soon!

Jack 的头像
Jack1 年前

Really nice find. The load balancer is a great idea

Faheem | FrontierMind AI 的头像
Faheem | FrontierMind AI1 年前

This dashboard looks super helpful for tracking usage across multiple Claude accounts. Load balancing between subscriptions and real-time monitoring of latency and costs would be great for managing budgets and debugging. Thanks for sharing!

Ariel Frischer 的头像
Ariel Frischer1 年前

This is cool - I also want to share which essentially for claude code power users who want to automate their workflows and run multiple agents in the cloud.

TreasureHunter 的头像
TreasureHunter1 年前

Does it work with bedrock as well

Ali Sherief 的头像
Ali Sherief1 年前

It's great but I don't think I have any utility for this as a Pro user.

MangoMagic™️ 的头像
MangoMagic™️1 年前

Track requests like a hawk. Transform data into your secret advantage.

Ashish Sheth 的头像
Ashish Sheth1 年前

Did I see $35 spend in last one hour? Is that correct? Definitely we need this tool.

YT 的头像
YT1 年前

is that possible to assign specific model for different sub agents through prompting?

Ian Nuttall 的头像
Ian Nuttall1 年前

I don’t think so yet!

Malcolm Little • Anatrova 的头像
Malcolm Little • Anatrova1 年前

I have a question for making a dashboard for a mobile app. Should I code it inside of the mobile app or create a whole new code base file that is a web app if that makes sense.

James Blackwell 的头像
James Blackwell1 年前

Noice

Prasenjit Sarkar 的头像
Prasenjit Sarkar1 年前

How about this collection of sub agents?

الحلم البعيد 的头像
الحلم البعيد1 年前

@grok "شرح هذا المنشور بالكامل وتوضيح المقصود منه بأسلوب مبسط ودقيق."

相关视频

how to use claude code mods like a top 1% user, step by step: give this to your agent before everyone catches on👇 1. set up Jev connect Jev to the model registry you want to use. check that the connection works and the listed models are available. 2. build your claude code mod open claude code 2.1.287 or later and paste this prompt: “build a mod called run-ledger. load plugin-authoring and use the API types for my installed version. create a dashboard that shows: - estimated cost per run, model, and source plugin where known - which model handles each task - each subagent’s status, latest action, and elapsed time include token counts, cache usage, and reported retries. count each request once. keep background tasks linked to their original run. use dated prices. label costs as API estimates, not subscription charges. show unknown when data is missing. connect the mod to my existing Jev setup. give Jev the task, available models, prices, budget, and relevant past results. ask it to recommend a model and explain why. start with recommendations. make automatic routing optional for eligible subagents. show the recommended model, actual model, result, and cost. include Jev’s own cost. keep state across hot reloads. add details and export. the dashboard itself must make no model calls. validate the plugin. test rendering, costs, attribution, and duplicate counting. give me steps for a live test.” 3. test the setup allow hot reload when prompted. run a simple task, a subagent task, and a mod-triggered model call. check that the dashboard updates, each request counts once, and Jev’s recommended model can actually run the task. 4. install the working mod ask claude to copy it out of the temporary folder and install it as a persistent plugin. see the cost. choose the model. check the result.

Avid

47,276 次观看 • 28 天前