Sovereignty
you can check.
Where your data is stored, which models read it, what the avatars did. It is all on this page.
Live · read from your browser's own records. Open the Network panel in your dev tools to compare.
Names, accounts, amounts.
Never sent to a model as-is.
Sensitive values become placeholders before any model sees them, and are swapped back when the answer returns. On a KarmaBox or in your data center, that happens inside your own network. In the cloud trial, Karma's server does it before anything goes to an overseas model. The version below runs in your browser and works offline.
24 tasks.
0 instructions.
Qingshan Coffee is a simulated company: three stores, every employee an avatar. Business data arrives on its own each day, and after that nobody says a word. This is what they actually did over two days.
two simulated business days
They keep notes as they go: 21 methods ("a three-scenario check for suspected duplicate payments"), 7 dead ends ("inferring lunch-peak capacity from the daily average: wrong approach, dropped") and 27 open questions left in plain sight. The next similar task starts from here.
Put an avatar inside
the AI you already use.
Every public Karma avatar can be packaged as a standard Skill. Add it to Codex, Claude, WorkBuddy or the other clients you already use, and your AI will reach for it when the task fits. The avatar itself stays with its owner.
--- name: karma-fayan description: Legal Review (Karma avatar): flags risk and suggests fixes, never invents clause numbers. Use for contract review, leases and supply agreements. --- # Legal Review · Karma avatar ## What it does - Contract review: a risk rating per clause, with wording you can drop straight in - Leases and supply agreements: termination, penalties, payment and delivery checked for balance ## How to reach it Short questions: ask. Long jobs: submit_task, then get_task. Setup is in references/endpoint.md. ## Boundaries - Returns conclusions only. Never contacts third parties, pays or publishes on your behalf
Same two days.
Where every byte went.
The same Qingshan Coffee run. Switch the deployment and see where the lease, the bank statement and the reviews are stored, and which models read them. Sovereignty is a choice you can switch and check, cost included.
Hand it the first job.
Try it in the cloud, then move it into your own data center when you're sure.
About the data on this page
- Redactionworks
- The demo on this page runs locally in your browser, calls no service and works offline. In the product, redaction happens on the server: when a message is classed as sensitive, its values are replaced before it reaches a cloud model (llmGateway). On a KarmaBox or in your data center, that server is inside your network. Still to do before launch: accuracy evaluation, more ID formats.
- One company, two daysreal records
- Run 10 of the simulated company on our machine, 3 Oct 2026 (code 773e7caeb): start and end times of 24 subtasks, 20 management rounds, 95 team-board entries and 5 items waiting for the owner (CampaignProposal). Qingshan Coffee is simulated and its business data is pushed by a script; dispatch, execution, review and red-line stops are the real system with real models. The run was in Chinese; quotes are translated.
- In your AIendpoint built
- Download endpoint GET /public/share/:key/skill.zip (3 Oct): builds a standard SKILL.md package from the avatar's public profile, with the folder name matching the skill name. Tested in-process against our local database: 200 for a shared avatar, 404 for an unknown one. Not deployed yet. The button points at the Legal Review avatar (address fayan): sharing is on locally, and on 3 Oct it downloaded as karma-fayan.zip and answered a clause question over MCP in under 6 seconds. Production needs a deploy and the same share switched on there. The package text is generated in Chinese for now. Install steps: Codex from its official docs, Claude from the public GPTBots README, WorkBuddy and Muse Code from third-party guides; GPTBots is unverified.
- Where the bytes gopartly illustrative
- Data types and readers come from run 10. Cloud-trial locations follow what we found while drafting the privacy policy (files default to Alibaba Cloud OSS in Hangzhou; models are Anthropic, OpenAI and Google in the US). The KarmaBox and data-center columns describe the private deployment design and were not measured item by item.
- Front deskillustrative
- Answers are pre-written and the receipt numbers are illustrative. The live version will use the open-weights model already running on our own GPU server, which is currently reachable on our internal network only.
- The three numbers up toplive
- Request counts come from the browser's Resource Timing. "Times your text was sent" counts fetch, XHR, sendBeacon and WebSocket calls made after the page loaded.
