Axiom / Private AI engineering

Code with AI.Keep it local.

An AI engineer that lives on your GPU. Axiom plans, builds and tests real software, remembers how you work, researches the web when you let it and renders images on your own card. Your code never leaves your machine, and every build ends in a verdict from tests it actually ran.

Windows 10 / 11 + WSL2NVIDIA GPU, 8 GB+Pre-launch — not on sale yet
Axiom / WorkspaceActual product
The Axiom workspace: asked to check the web for the latest Python release, Axiom answers from 8 live official sources with numbered citations, beside the project's memory and code graph.
A live web answer, sources cited, beside memory and the code graph.View full size
01 / The workspace

Everything the job needs.
In one window.

Your files, your models and the evidence behind every change sit side by side, so you can see what the AI knows before it touches anything.

01 / Context

It knows your repo.

Axiom maps your files, functions and tests into a project graph, and carries project memory from one session to the next.

Files / Graph / Memory
02 / Control

Nothing moves without you.

Axiom lays out its plan and waits for your approval before it builds. Generated code runs inside a sandbox.

Plan / Approve / Sandbox
03 / Verification

It shows its work.

Every build runs real tests. Red names the test that failed, and anything only checked by its own tests is labelled unverified.

Test / Verdict / Evidence
Axiom / One real build, start to finishActual product
Asked for a trip expense splitter, Axiom wrote a plan, waited for approval, built and tested it in the sandbox (green: all 9 of its own tests pass), then ran it in the preview. Alex paid $120 for dinner, Sam $60 for gas, Priya $45 for tickets: Priya pays Alex $30, Sam pays Alex $15. Recorded on a 32 GB GPU; the ten-minute build plays 32× faster.
02 / Capabilities

It remembers.
It researches. It renders.

Long-term memory, live web research and image generation are built in, and they run on hardware you already own. You decide what's switched on.

Memory · on by default

Never explain twice.

Tell Axiom a rule once and later chats and builds follow it. Lessons from past fixes flow into new ones: on an 8 GB card, repairs dodged a trap they had hit before 8 times in 10 in a pre-registered live test. It all lives as plain Markdown in a git-backed vault on your disk.

Facts / Lessons / Vault / Forget anytime
Web · off until you switch it on

Answers with receipts.

Switch web on and ask about anything newer than your model: release notes, API changes, today's docs. Axiom searches, reads the pages and answers with a numbered link to every source. Deep research works one topic across many pages. Every fetch is guarded and logged, and one switch cuts the line again.

Live search / Citations / Deep research
Images · installs on your click

Pictures from a sentence.

Ask for an image and Axiom renders it with ComfyUI on your own GPU. No credits, no uploads. It picks a model that fits your card, from DreamShaper on 8 GB to Juggernaut XL and FLUX.1 on bigger ones, and writes its own ComfyUI workflows, checked against your install before they run.

ComfyUI / SDXL / FLUX.1 / Local renders
03 / Proof facts

Measured.
Not promised.

Every number below comes from a recorded run on real hardware, not a spec sheet.

Measured by Vantyrx Labs on Axiom 8000-series builds, October 2026. Verdicts were graded against hidden tests Axiom never saw.

Every part runs
on your machine.

Local models on your GPU do the thinking. Axiom's kernel owns the disk and the sandbox. Nothing in this loop needs the internet.

How Axiom runs a build on your machine Inside your machine: you send a request from the app, which talks to the engine over HTTPS on 127.0.0.1 port 8766. The front model reads it with context from the project graph and memory and plans. You approve the plan. The coder model writes a patch, and the kernel's write path puts it on disk. The sandbox, bubblewrap with no network, runs the tests; a red result loops back to the coder for a repair. The verdict, with what passed, what failed and what is unverified, comes back to you as a proof card. Both models run through Ollama on 127.0.0.1 port 11434 on your GPU. The sandbox's link to the internet is cut, and with web left off, the default, Axiom makes zero outbound connections. Your machine · Windows 10 / 11 + WSL2 · loopback only promptplangopatchresultsreport write red → repair context 01 / ui :8766Youask in plain words 02 / front agentFront modelreads + plans 03 / gateApproveyour call 04 / coderCoder modelwrites the code 05 / write pathKernelowns the disk 06 / bwrap no-netSandboxruns the tests 07 / proof cardVerdictpass · fail · unverified 08 / memory planeProject graph + memoryfiles, functions, tests, past work Local models on your GPU · 8 / 16 / 32 GB VRAM 0 outbound · web off ext / wanInternetoff by default legend data flow repair loop context axiom / build looprev 2026.10 · 1/1
A figure made of light and code holds a finger to her lips.
Axiom / Local and private
04 / Built on boundaries

Your machine.
Your rules.

Local first.

Your code, prompts and files stay on your machine. Web and cloud models stay off until you switch them on.

Controlled changes.

Edits pass through Axiom's own write path, not straight from a model to your files.

Pre-launch

Get in early.
Not on sale yet.

Axiom isn't on sale yet: checkout is closed. Leave your email and your GPU size, and we'll tell you the day it goes on sale.

Pre-launch · Not on sale yet

Axiom Local

$250 / year

The planned price for a year of private, local AI engineering.

Your license is signed and checked on your own machine, and the engine never phones home on its own. Work fully offline for up to 60 days between license check-ins; installing a signed update or re-activating renews the window.

  • No per-token bills
  • 60 days offline
  • 12-month term
  • One machine at a time
Not yetCheckout is closed. Axiom is not on sale today.

Join the early-access list

You're on the list. We'll email you the day Axiom goes on sale.

We keep your email and GPU size for one thing: telling you when Axiom launches. No sharing, no other mail. Ask to be removed any time.

One active machine at a time; a key can be moved up to 5 times per 30 days. The price and the terms above are the plan for launch, not an offer you can accept today — joining the list costs nothing and commits you to nothing.