Used up its entire 2026 AI budget in four months.
After giving Claude Code to about 5,000 engineers, the cost per engineer hit $500 to $2,000 a month.
Source: Fortune →Clawgate sits between your developers and AI providers, giving engineering teams complete visibility, governance, and cost control across AI coding tools like Claude Code, Codex, and OpenCode.
Monitor usage. Enforce budgets. Reduce token usage. Optimize model selection. Track every AI dollar by project. The savings AI promised, without the surprise cost.
The question every leader faces
Do you know what your AI coding tools are actually costing you?
Can you see who's using AI, where it's being used, and whether you're overspending?
Clawgate answers all
fewer tokens sent on heavy workloads
to point your CLI at Clawgate
over budget — caps are hard stops
tokens processed through Clawgate
When you give AI to your team, you can’t see the spend until the invoice arrives. The biggest companies in the world are finding this out the hard way.
Used up its entire 2026 AI budget in four months.
After giving Claude Code to about 5,000 engineers, the cost per engineer hit $500 to $2,000 a month.
Source: Fortune →
Moved to cap employee AI usage as costs approached billions.
Staff burned 73.7 trillion tokens in 30 days, racing to top an internal leaderboard.
Source: The Information →Company AI runs on private side projects. You still pay.
Which project cost what? It’s one big number.
The CLI picks the model. You pay for its choice.
Every request is checked against your rules before it runs — not reported back to you after the invoice arrives.
Daily and weekly caps per user and per project. Hit the cap and the request stops.
You pick what your team can run — or let smart routing serve the cheapest model that fits each request.
Every dollar tied to a developer and a project, in real time.
Up to 92% fewer tokens on heavy workloads, with answers unchanged.
See the numbers →Built at Virstack to control our own AI bill. It cut it by 70% before it was a product.
Most requests don’t need your most expensive model. Clawgate sizes up each one and serves the cheapest model that can handle it.
Pick the range you’re happy to pay for — a frontier model at the top, a low-cost one at the bottom.
Each request is scored on the quality it needs, what it would cost, and how fast it should come back.
Routine work runs on the cheap model. Your top model is saved for the hard parts. One switch, no tuning.
AI makes your team faster either way. The question is whether the cost climbs with the output, or races past it.
The cost blows past the budget
The cost stays inside the budget
With Clawgate, the gains from AI productivity stay in your business, instead of leaking out as ungoverned spend.
Illustrative model. A 200-developer org at a mid-range $800 / dev / month of uncontrolled AI spend.
This is an illustrative model, so the shape matters more than any single figure. It assumes a team of about 30 developers using Claude Code, Codex, and OpenCode daily, with per-developer spend anchored to the $500 to $2,000 per engineer each month that companies like Uber reported above. The budget grows with output. Left ungoverned, spend drifts above it as always-on top-tier models, retries, and side projects pile up, while governed spend stays inside the cap for the same output.
These aren't numbers we invented for a slide. Clawgate started as an internal tool at Virstack LLC: we built it to keep our own engineers' AI spend under control before we offered it to anyone else. The pattern above is the one we watched play out on our own spend first.
One vsk_…
key. Billed at cost, plus one clear platform fee.
One endpoint change for your team. After that it’s dashboard toggles: set a budget, pick the models, watch the spend.
Claude Code, OpenAI Codex, OpenCode, and Cursor. Your developers keep the tool they already use — Clawgate sits behind it as the gateway those tools talk to, so nothing about their workflow changes.
No. Setup is a one-line endpoint change plus a vsk_… key per developer, and a single command writes that config for them. There is no new tool to learn, no plugin to install, and no change to how they prompt. See how it works →
It is a hard stop. Budgets are checked before a request is sent, so once a daily or weekly cap on tokens, dollars, or sessions is reached, the request is refused with a clear message rather than billed. You can also set a warning threshold — say 80% — so the key owner gets an email before anyone hits the wall.
Yes. Set an allow-list per team or per project, or pin everyone to a single model. Ask for a model outside the list and Clawgate serves the best one you have allowed instead of failing the request. Turn on smart routing and routine work is served by a lighter, cheaper model automatically. See the model catalog →
We do not store your raw prompts or source code, and we never use your request content to train machine-learning models. What we record per request is usage metadata: the model used, token counts, computed cost, latency, and timestamp, attributed to a user and project. Abuse detection works from privacy-preserving, non-reversible fingerprints rather than the content itself. Read the privacy policy →
Minutes. Create an organization, issue a key per developer, and run one command on each machine. Budgets and model rules are dashboard toggles after that — no code and no config files to maintain.
Yes, on the Enterprise plan. Clawgate runs inside your own infrastructure, usage data stays in your environment, and there is no per-token markup — you pay the provider directly. Talk to us about Enterprise →
Give your team the AI they need. Give yourself the visibility and control you need.
Anthropic
OpenAI
xAI
DeepSeek
Moonshot
Z.ai
Sakana AI
NVIDIA