<aside> 🔏
Your best work is the work you can't paste into a chatbot. Kakurega is the private AI workspace that proves what it did. Chat, documents, meeting notes, memory, a team of experts and a full code review, on your own computer. Keys and passwords are withheld before a model sees them, every finding is reproduced in a sandbox, nothing leaves until you approve the exact bytes, and what leaves is a signed capsule anyone can check offline. In the cloud, the model runs in confidential hardware on NEAR AI Cloud, with its proof checked before every question.
Live site · GitHub · Demo video · Watch a real review · Check a capsule · Ask NEAR yourself · Try the gate
</aside>

A review run on NEAR AI Cloud: a finding the kernel confirmed, with the script it ran and what the script printed
<aside> 🎯
0 of 126 known secret shapes reached the model (signed benchmark)
</aside>
<aside> ✅
7 of 7 no-key judge checks pass on a fresh Linux clone
</aside>
<aside> ⛓️
20 model signatures of one real review verified on NEAR testnet
</aside>
<aside> 📑
267 claims in the public docs checked against the code by a script
</aside>
One workspace for everything you do with AI, where a security layer outside the model decides what may leave your computer. An agent reads your project inside a sealed sandbox with no network. Keys and passwords the detector knows are withheld first. Each defect it reports comes with a script, and the kernel reruns that script before you ever see the finding. You approve the exact files and their digest outside the page. What leaves is one signed file the recipient verifies offline. When the model is not on your computer, it runs in confidential hardware on NEAR AI Cloud: the hardware proof is checked before a question may go, and every answer carries the model's own signature.

The path of a result: your files, the sealed sandbox, the gate outside the model, what is withheld, held for your yes and leaves signed; below, where the model runs and what is on NEAR testnet
For users
If you work with private material
If you are an engineer or security reviewer

The sealed room of a real review: one dot per file. The red one was withheld and never reached the model.
<aside> 🪤
Try to break it yourself: the gate on the home page runs the product's own secret detector in your browser on your own text, and nothing is sent. The live try page goes one step further: it withholds the keys, then sends the rest, encrypted, to a model in confidential hardware on NEAR AI Cloud and checks the signed answer. In the app, "Try to break it" plants your own attack next to two fresh secrets and shows exactly what the model was sent.
</aside>
For judges
<aside> ✅
The short version: a working prototype you can check yourself, not a deck. One command runs seven checks with no key, no account and no funds. A live site replays a real review from its capsule and lets you break that capsule. Every limit is written next to the claim it limits.
</aside>
git clone https://github.com/xaorao/kakurega && cd kakurega
scripts/judge
Expected: seven lines starting PASS, then 7 passed, 0 failed, 0 skipped. (needs Rust; the first build takes several minutes).
| Working product | Browser workspace, desktop app, terminal workspace, extension, Telegram bridge and MCP server on one Rust kernel. Every feature and its state |
|---|---|
| Hosted demo | kakurega-eight.vercel.app: replay of a real review, capsule check, NEAR proof page, the gate on your own text |
| Reproducible README | "Do X, see Y" steps with expected output; docs/JUDGES.md; a repo-root llms.txt mapping each claim to its proof |
| Tests | Unit, integration and end-to-end suites across every surface, sandbox suites on Linux, hostile cases and mutation checks for the safeguards; CI green on GitHub |
| Honesty | docs/LIMITS.md: what is not verified, known weaknesses, what is deliberately not built |
| Licence and credits | Apache-2.0; every library listed in NOTICE.md; the README says most code was written with AI coding assistants |
| Privacy & Private AI (primary) | Confidential inference on NEAR AI Cloud with the hardware proof checked before every question and a signed receipt under every answer; a model on your own computer as an equal route; secrets withheld; "Where your data goes" read from the signed log. |
|---|---|
| AI Security & Defense (second, a partial fit) | A policy gate outside the model, a sandbox with no network, planted instructions passed as data, approval of exact bytes outside the page, leases for outside agents. Named honestly as partial: that track centres on giving AI money safely; ours protects data. |
| NEAR AI Cloud | The private cloud route: Intel TDX and NVIDIA evidence checked against a fresh challenge, questions encrypted to the model, every answer's signature checked before it is shown. Live with Qwen/Qwen3.8-27B. |
|---|---|
| NEAR's on-chain verifier | A registry contract hands the model's quote to NEAR's own verifier contract and records the key it vouched for. Live on testnet, readable by anyone. |
| Capsule registry and orders | Who registered a capsule, and that its model signatures verified. An order holds payment for a review and pays only when the named reviewer registers a capsule for it. Live on testnet. |
Remove NEAR and the private cloud route goes, and with it the public record of the hardware check that a recipient can read without us.
<aside> ⚠️
Research
PMF
Capability hub
Open items
| --- | --- | --- |
Every picture is from a real run on NEAR AI Cloud.