I got tired of every AI tool requiring a subscription and storing conversations on someone else's server, so I built Axiom — a Windows desktop app that runs entirely on your machine.
The main feature is the council pipeline. Instead of one model answering your question, three independent models work sequentially. The Architect breaks down the problem into a numbered plan. The Builder implements it. The Critic reviews the output and flags issues before you ever see it. If the Critic finds problems, the Builder revises. You get a confidence label on every response High Confidence, Moderate Confidence, or Flagged for Review.
It also has a normal chat mode with a Python sandbox that actually executes code and verifies math results rather than guessing, web search that extracts real page content not just snippets, LaTeX rendering for equations, and optional cloud mode via OpenRouter if you want access to larger models like GPT-OSS-120B using your own free API key.
The default model is Volta, based on Qwen3-4B Q4_K_M — about 2.5GB, runs on any modern PC without a GPU.
Everything is free. MIT licensed. No telemetry. Nothing leaves your machine in local mode.
Download and source on GitHub: [https://github.com/YoMosa2009/Axiom]
Happy to answer questions about how the pipeline works.
Top comments (1)
The council pipeline is the interesting bet, because three independent models debating is really a verification strategy in disguise: a single model has no way to catch its own confident-wrong answer, but a second and third with different priors can surface the disagreement that flags it. That's the strongest argument for multi-model, you're not buying more intelligence, you're buying a disagreement signal, and disagreement is exactly where you don't want to trust the first answer. Two design questions decide whether it pays off. First, diversity: three runs of the same model mostly agree with each other's mistakes, so the council only works if the models genuinely differ (different families/training), otherwise it's expensive redundancy. Second, the cost-vs-task tradeoff: 3x inference for every question is a lot, so the council earns its keep on high-stakes or ambiguous queries and is overkill for the easy ones, ideally you escalate to the full council only when a cheap first pass shows low confidence. Local-and-free changes that math nicely, since the 3x is compute you already own, not 3x a bill. That use-disagreement-as-verification-but-route-by-stakes instinct is core to how I think in Moonshift. Do all three always run, or do you short-circuit when the first two agree confidently?