New Axiom AI for Windows

The same New Axiom AI in your terminal — streamed answers with their proof tier, the exact-gate list, your account, code mode and Autonomous Work. Python 3.9+:

pip install https://newaxiomai.com/web/dl/newaxiom_ai-0.3.0-py3-none-any.whl

newaxiom ask "is 1009 prime?"   # works immediately (guest workspace)
newaxiom login                  # use your account
newaxiom gates                  # the exact-gate reality layer

Accounts are created on this site (email verification), then newaxiom login in the terminal. The gates and models run server-side — the client only renders what the server actually claimed.

Code mode

Point it at a project and it reads the repository, proposes a diff, and lets you approve it file by file:

cd your-project
newaxiom code "add a --json flag and a test for it"

--dry-run            show the diff, write nothing
--test "pytest -q"   run YOUR tests before and after, and report the difference
--local-key          use your own OPENAI_API_KEY; this site is not contacted
--undo               put the last run back, byte for byte

What leaves your machine. Code mode lists every file it would send and asks once per folder before sending anything. Files that look like credentials — .env, *.pem, id_rsa, anything containing a key-shaped string — are withheld and named to you, and .git/, .venv/ and node_modules/ are never read. It needs an account, because the shared guest workspace is not a place for your code. With --local-key nothing reaches this site at all. Details in the privacy policy.

What it writes. Only files you approved, atomically, preserving your line endings and BOM, with the whole batch rolled back if any single write fails — and a snapshot taken outside your repository first, so newaxiom code --undo always has something to restore. Python and JSON files are parsed before you are asked; everything else is reported as unchecked, per file. Parsing is not testing, and this lane never runs a command you did not give it.

Autonomous Work

Point agents at your own chats, bound the run in time and money, and read back what they explored — from the terminal, and without holding it open:

newaxiom chats                                   # your chats, with their ids
newaxiom work start --chat c_ab12 --minutes 15 --agents 4 --follow
newaxiom work list                               # your runs + what your plan allows
newaxiom work show                               # the last run in full, tiers and all
newaxiom work cancel                             # ask a run to stop

Runs live on the server, so --follow is a view, not a leash: close the terminal and the agents keep going. Time, agent count and model spend are capped by your plan, and a cap that changes what you asked for is printed rather than applied quietly. Every finding carries the tier the server recorded — agent proposals are heuristic; only an exact gate receipt makes one verified.