Isolated by default
A project declares where its work happens. The default is a disposable container mounted only at that project: no packages installed on the laptop, nothing left in the home directory.
Winnex AI
Ask it. Then look at what it cost.
This is the agent that runs on your machines, answering questions about itself over a corpus you can see. Every answer arrives with the cost of that exact query and a hash of the record — because an AI you cannot audit is an AI you cannot put in front of a customer.
Read-only mode. The agent reads a public corpus, searches it and answers. It cannot run commands or write files.
What you just saw
The numbers under the answer are the product.
A typical question here costs a few hundredths of a cent. Measured across nine developer tasks, this agent used 2,097 input tokens where the alternative used 6,379 for the same work — and that gap does not shrink as tasks get bigger.
For a company the saving is real but it is not the reason to buy. The reason is the line under the answer: the cost, the time, and the hash of the record. When someone asks in six months what the AI did, that line is the answer.
For companies
It runs on your machines, on your key.
The agent is a 1.8 MB binary. Your developers run it where the code is; the model endpoint is the one you already pay for. Nothing of yours passes through us — not a prompt, not a file, not a result.
Auditable by construction
Every tool call becomes a block in a SHA3-256 hash chain, append-only, outside the working directory. It proves the record was not edited after the fact, and it survives the container being thrown away.
One person, a thousand machines
A controller key, an inventory in JSON, and one command applied to the fleet in parallel. Failures are listed first and named by machine.
Next step
Run this on your own repository.
Tell us the size of the team and we will send a pilot key for thirty days, the entitlement flow, and the audit format your compliance team will want to read.