Control your team's AI spend on Claude Code and Codex with hard-stop budgets, model controls, and per-project cost visibility. Built in prompt compression cuts token spend by up to 92% on heavy agent workloads. Never a surprise bill.
Viktor.comAn AI coworker that actually does the work
Promoted
Maker
📌
Hey Product Hunt 👋
I'm CK, Engineering Lead at Virstack LLC. Clawgate started as an internal tool we built to solve our own problem.
Here's what happened. Like a lot of teams, we gave our engineers Claude Code and Codex, and the productivity jump was real and immediate. But a few weeks in, our AI bill had climbed in a way none of us could explain. We couldn't see who was spending what, which projects were burning the most, or why a documentation task had quietly run on a frontier model all weekend.
Turns out we weren't alone. Uber burned through its entire annual AI budget in four months. Meta's engineers ran up 73.7 trillion tokens in a single month racing an internal usage leaderboard. If it can happen to them, it can happen to a 30-person team. It happened to us.
So we built Clawgate. It's an API gateway and dashboard that sits between your developers and the AI providers.
What it does: 🔹 Track spend by user and by project, in real time 🔹 Set hard budget caps (daily/weekly, in sessions, tokens, or dollars) that actually stop the bill
🔹 Get usage alarms and email notifications when spend crosses a threshold you set (fully configurable) 🔹 Control which models each team can use & keep the expensive ones for who truly needs them 🔹 Mix and match models across providers, so Claude Code and Codex run on whatever's cheapest for the job 🔹 Compress bulky context before it hits the model (up to 92% fewer tokens, accuracy held) 🔹 Catch abuse: leaked keys, spikes, shared credentials without ever reading your code
Setup takes minutes. It's a one-line endpoint change in Claude Code or Codex, and your team never has to learn a new tool or change how they work.
The whole idea: keep the speed AI gave you, without the surprise bill.
We'd genuinely love your feedback! What would make this a no-brainer for your team? And if you're running Claude Code or Codex across a team right now, I'm especially curious how you're handling cost today. Happy to answer anything in the comments 🙏
Report
The Slack alerts for budget thresholds would be way more useful if they also showed which user or session triggered the spike, so we can ping the right person instead of hunting through logs.
Report
Maker
@elifeaykarep1p Yes, this is exactly right, pointing you straight at the user or session is the difference between a useful alert and more log-digging. Email alarms are live and configurable today, Slack's on the way, and we'll bake the who-and-where into it. Thanks for flagging.🙏
Report
Hard budgets are great, but adding a Slack or email alert at like 80 percent of the limit would be clutch. Catching it before it slams the door means I can decide whether to top up, swap to a cheaper model, or just pause that project for the day instead of scrambling.
Report
Maker
@samitbqn Good news, that's already in, you can set configurable usage alarms with email notifications at whatever threshold you want, so 80 percent is exactly the kind of early warning it's built for. Slack alerts are on our radar to add next. Appreciate the thinking here, it's exactly how we hoped people would use it.
Report
Caught my attention because the prompt compression actually moved my token numbers noticeably on a long agent run, not just a vague claim. Budget caps per project are the part I wish more tools had.
Report
Maker
@etintavukur1vd Appreciate this 🙏, the compression showing up on real runs is exactly what we were after. Per-project caps were born from our own pain too. Let me know how it goes on bigger workloads.
Report
honestly the prompt compression claim is what caught my eye, 92% on heavy agent workloads is a wild number and basically the reason I'd even consider switching tools. nice execution on something that's been a real pain point
Report
Maker
@serhatkulaber That's the one we're proudest of, and also the one we knew people would rightly be skeptical of, so it's measured on real agent workloads. The heavier the run, the more stale context there is to compress, which is why the number lands where it does. Give it a try and tell us what you get. 🙏
Report
finally something that stops me from getting blindsided by a runaway agent run at 3am. the per-project visibility is super clear and the prompt compression actually moved the needle on my heaviest workflows
Report
Maker
@devran1236463 The 3am runaway is exactly the nightmare we built the hard caps for, so glad you're covered now. And the compression moving the needle on your heaviest workflows is what we were after, not just a number on a slide. Appreciate your comment. 🙏
Report
finally something that actually puts a hard ceiling on claude code spend, the per-project cost breakdown showed me exactly where tokens were bleeding. prompt compression alone saved us a noticeable chunk on heavier workflows
Report
Maker
@zzet1032068 Love hearing this, the per-project breakdown showing exactly where tokens bleed is the whole reason we built it. And the compression saving you real money on heavy workflows is exactly what we were after, not a vague claim. Thanks for commenting. 🙏
Hey Product Hunt 👋
I'm CK, Engineering Lead at Virstack LLC. Clawgate started as an internal tool we built to solve our own problem.
Here's what happened. Like a lot of teams, we gave our engineers Claude Code and Codex, and the productivity jump was real and immediate. But a few weeks in, our AI bill had climbed in a way none of us could explain. We couldn't see who was spending what, which projects were burning the most, or why a documentation task had quietly run on a frontier model all weekend.
Turns out we weren't alone. Uber burned through its entire annual AI budget in four months. Meta's engineers ran up 73.7 trillion tokens in a single month racing an internal usage leaderboard. If it can happen to them, it can happen to a 30-person team. It happened to us.
So we built Clawgate. It's an API gateway and dashboard that sits between your developers and the AI providers.
What it does:
🔹 Track spend by user and by project, in real time
🔹 Set hard budget caps (daily/weekly, in sessions, tokens, or dollars) that actually stop the bill
🔹 Get usage alarms and email notifications when spend crosses a threshold you set (fully configurable)
🔹 Control which models each team can use & keep the expensive ones for who truly needs them
🔹 Mix and match models across providers, so Claude Code and Codex run on whatever's cheapest for the job
🔹 Compress bulky context before it hits the model (up to 92% fewer tokens, accuracy held)
🔹 Catch abuse: leaked keys, spikes, shared credentials without ever reading your code
Setup takes minutes. It's a one-line endpoint change in Claude Code or Codex, and your team never has to learn a new tool or change how they work.
The whole idea: keep the speed AI gave you, without the surprise bill.
We'd genuinely love your feedback! What would make this a no-brainer for your team? And if you're running Claude Code or Codex across a team right now, I'm especially curious how you're handling cost today. Happy to answer anything in the comments 🙏
The Slack alerts for budget thresholds would be way more useful if they also showed which user or session triggered the spike, so we can ping the right person instead of hunting through logs.
@elifeaykarep1p Yes, this is exactly right, pointing you straight at the user or session is the difference between a useful alert and more log-digging. Email alarms are live and configurable today, Slack's on the way, and we'll bake the who-and-where into it. Thanks for flagging.🙏
Hard budgets are great, but adding a Slack or email alert at like 80 percent of the limit would be clutch. Catching it before it slams the door means I can decide whether to top up, swap to a cheaper model, or just pause that project for the day instead of scrambling.
@samitbqn Good news, that's already in, you can set configurable usage alarms with email notifications at whatever threshold you want, so 80 percent is exactly the kind of early warning it's built for. Slack alerts are on our radar to add next. Appreciate the thinking here, it's exactly how we hoped people would use it.
Caught my attention because the prompt compression actually moved my token numbers noticeably on a long agent run, not just a vague claim. Budget caps per project are the part I wish more tools had.
@etintavukur1vd Appreciate this 🙏, the compression showing up on real runs is exactly what we were after. Per-project caps were born from our own pain too. Let me know how it goes on bigger workloads.
honestly the prompt compression claim is what caught my eye, 92% on heavy agent workloads is a wild number and basically the reason I'd even consider switching tools. nice execution on something that's been a real pain point
@serhatkulaber That's the one we're proudest of, and also the one we knew people would rightly be skeptical of, so it's measured on real agent workloads. The heavier the run, the more stale context there is to compress, which is why the number lands where it does. Give it a try and tell us what you get. 🙏
finally something that stops me from getting blindsided by a runaway agent run at 3am. the per-project visibility is super clear and the prompt compression actually moved the needle on my heaviest workflows
@devran1236463 The 3am runaway is exactly the nightmare we built the hard caps for, so glad you're covered now. And the compression moving the needle on your heaviest workflows is what we were after, not just a number on a slide. Appreciate your comment. 🙏
finally something that actually puts a hard ceiling on claude code spend, the per-project cost breakdown showed me exactly where tokens were bleeding. prompt compression alone saved us a noticeable chunk on heavier workflows
@zzet1032068 Love hearing this, the per-project breakdown showing exactly where tokens bleed is the whole reason we built it. And the compression saving you real money on heavy workflows is exactly what we were after, not a vague claim. Thanks for commenting. 🙏