Why we built this
Manus is a great product. The UX is the right shape for non-coders: a single intent bar, a live task card, files you can download at the end. But it is priced like a luxury good. A task that costs $0.30 in raw Claude inference is billed at roughly three times that. We thought there was room for a clone that kept the UX and dropped the markup.
Pass-through pricing
AtmosphereAGI bills at Anthropic's sticker rate. Sonnet 4.6 input is $3 per million tokens, output is $15 per million. We pass-through with zero markup. Your invoice line-items each step at the same rate Anthropic would charge you directly.
How we make money
We route the cheap parts — planning, summarisation, classification — to local models we run on our own GPU. Result chunks get summarised by our caveman-mode skill before being sent back into Claude, which cuts output tokens by 70–80%. Prompt caching, batch mode, and replay-the-known-good-trajectory finish the savings stack. The margin lives in our infra, not in your invoice.
The math
Manus 'Pro' task ~ 50k in + 8k out
raw = 50k * $3/M + 8k * $15/M = $0.27
Manus = ~$0.80 (3x markup)
us = $0.27 (sticker, plus our infra-side savings)BYO key, if you want
If you already have an Anthropic key, plug it in. Your key is encrypted with AES-256-GCM at rest, only the last four characters are ever shown back to you, and deletion is immediate and irreversible. With your key, you pay Anthropic directly for inference and us a small flat fee for the runtime.
What is here today
Eight seed agents (email, research, slides, web worker, voice notes, logo, code, numbers), a live three-panel task page, real-time spend counter, scheduled tasks, history with replay, public share links for finished runs, and a billing portal. The email agent is free forever. Sign up takes under a minute.
— the cave crew