A small open model, like Gemma 4, runs on the device. Completely offline. No warehouse in the middle.
CIRIS · Open Coherence Infrastructure
One app. No data centers. Every claim it makes, signed.
CIRISAgent · CEWP “soup”CIRIS 2.9.4 shipping · 3.0 rolling out June and July 2026
Where we are
Ask an AI anything and your words travel to someone else’s building, run on someone else’s model, and land in someone else’s logs, with the carbon as a side effect.
You don’t own the answer, the data, or the footprint.
The shape of the alternative
CIRISAgent is what you hold and talk to. It carries the reasoning.
CEWP is the fabric it runs on: a web of devices that trust each other, all hardware you already own.
People and AI both live on it as equals, side by side.
CIRISAgent
the app · reasoning + UI
~14 KB trace / decision
CEWP “soup”
the fabric · your hardware
The big idea
Usually these come from three different companies. Here they fall out of one design.
01
It runs on the phone in your pocket and a small box at home.
02
Your private life literally can’t leave your circle.
03
Every claim it makes is signed, so you can see it and undo it.
PROMISE 01
65%
of everyday web activity never leaves your city. We call it the locality bonus.
phones · home boxes · no giant data centers
Two paths · same signed agent
A small open model, like Gemma 4, runs on the device. Completely offline. No warehouse in the middle.
For phones too small to run a model, CIRIS is free over hosted inference in 29 languages. Same signed claims. Someone else’s data center.
PROMISE 02
selfglobal commons
The agent sorts everything into seven circles, from self out to the whole Global Commons.
For the closest circles, the system never sends the little message that tells the network your data is even there. No message, no way to ask for it. The privacy comes from what’s missing, not from a promise.
PROMISE 03
Big tech checks its AI once, while building it, then hopes. CIRIS keeps checking it while it runs, out in the open.
When the AI gets something wrong, you don’t just see it, you can undo it, and it loses trust.
Claim
"Here's the safest dose for that."
signed · ed25519+pq · 0xA17F…9C2
And here’s the safety net
"This is safe to take."
"That report is solid."
"This source checks out."
every claim → moot
If a person, or an AI, is ever caught misleading people, every claim they ever signed becomes moot, and every conclusion built on them can be re-opened.
Why it’s one system, not three
every byte is signed
who made it · whether it can leave
No data centers
AI you can check
Privacy that can't leak
The same signature that proves who said something also decides what stays home. One simple idea, and privacy, less waste, and safe AI all come from it.
And we’ll say it out loud
It works before the AI can even think, and only a trusted, signed key can set it off.
Every choice is signed and linked to the one before. A lie can’t keep its story straight.
Checked by computer on every update, built first for the people who need it most.
And every word of it is open code you can read.
Not just talk
Open source, all the way through. 2.9.4 is shipping now on iPhone and Android, and 3.0 rolls out through June and July 2026.
github.com/CIRISAI · AGPL-3.0 · mission-lockedAn AI you can own
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。