Axiom — AxiomInference

One workspace that adapts to the model you give it.

Axiom measures a local model's parameter count and scales context budgets, tool routing and prompt complexity to match — so a 0.5B model and a 70B model both produce usable results instead of the small one collapsing.

01

Local-first by default

Chats, API keys, connector tokens, models and Workplace sessions live in %LOCALAPPDATA%\Axiom and are never placed inside release packages.

02

Chat & Workplace

Everyday chat with every tool available, plus a multi-agent Workplace for work that takes more than one pass.

03

Tools that actually run

Python and Java sandboxes, a calculator, multi-source web search with trust scoring, and controlled codebase access.

04

Updates itself

Checks the stable GitHub Releases feed at startup, verifies the package, and restarts in place — leaving your chats, models and settings untouched.

Three ways to run a model.

All three are first-class. Attachments, Skills, tools, Project Canvas and memory behave the same way whichever one is selected.

	Runs on	You need	Good for
Local	Your GPU/CPU via llama.cpp	A GGUF model	Full privacy, offline work, no per-token cost
Hybrid Local	Your own server	An OpenAI-compatible endpoint	Your own hardware, reachable from anywhere — this is how Axiom talks to a self-hosted Kestrel 1
Cloud	OpenRouter	Your own API key	Frontier-class models when you want them

Workplace

A multi-agent workspace for work that takes more than one pass.

Council mode runs three roles in sequence: an Architect plans the task, a Builder produces the deliverable, and a Critic reviews it and can send it back for a targeted revision.

Single Model mode collapses those handoffs into one agent that plans, uses tools, executes and verifies its own result. Switch between them from the Workplace header whenever a run isn't active.

Persistent sessions, task history and per-role context controls
Live activity and completion notifications
Project Canvas: type @ProjectCanvas and ask for a file — a requirements.txt, a docker-compose.yml, a CSV — and it appears under its real name in a resizable side pane. CSV renders as a table, Markdown renders, HTML and SVG keep their preview.

Axiom's Workplace tab, where Council or Single Model runs are set up. — **Workplace.** Council and Single Model workflows, with the Project Canvas the Builder writes into.

Agents that can act on your machine — on your terms.

Two opt-in Workplace features let a model operate the computer directly. You choose how much rope each one gets.

Computer Agent

Turn it on in the Workplace sidebar and the model works the way a terminal coding agent does — running commands, reading and writing files, searching the disk — then answers from what it actually found.

It has eight tools: run_command, read_file, write_file, append_file, edit_file, list_directory, find_files and search_text. Before a run may finish, Axiom checks the claimed result against the disk: a summary listing files that were never written is sent back to the agent to fix.

Mode	Behaviour
Manual	Stops before every command and file change: Approve, Always allow this, or Deny.
Auto	Runs without stopping to ask.

Always refused, in both modes: formatting a drive, repartitioning, deleting a drive root, rewriting boot configuration, wiping free space, shutting the machine down, and writing into Windows system folders. The patterns are narrow — rm -rf ./build and git reset --hard run normally. Whole-computer scope and Auto mode each need an explicit confirmation.

Computer Use

Type @ComputerUse followed by a goal in the Workplace composer and Axiom drives the desktop directly: screen capture, planning, mouse and keyboard, and verification that the action actually landed.

Workplace composer

@ComputerUse open Notepad and type hello

It needs a vision-capable model (a cloud vision model, or a local GGUF with a projector). Axiom probes the active model first and tells you plainly why a run can't start rather than failing halfway. Sessions run in a dedicated window with a visible pointer overlay, an action log, and safety checks on navigation and target selection.

Computer Agent works on Local, Hybrid Local and Cloud. Capable models get JSON tool calls and a long step budget; sub-4B models get a flat text protocol they can actually produce, so the feature still works on a small local model.

Axiom's Skills panel listing installed Skills. — **Skills panel.** Attach a Skill once and it applies everywhere.

Skills & Plugins

Reusable procedures that deliver real artifacts.

Skills apply across Local, Cloud, Hybrid Local, Council and Single Model. Some answer in chat; others render a finished artifact in Project Canvas.

Skill	Delivers
Slide Deck Studio	A navigable 16:9 deck with arrow-key navigation and speaker notes
PDF Studio	A print-ready document with a proper print layout
Data Analysis	A charted report with inline SVG charts and the figures beneath them
Document Summarizer	A traceable summary in chat
Code Review	A review in chat

It works on small models too. A sub-4B model can't reliably author a self-contained HTML document, so below that threshold the model supplies a short structured outline and Axiom builds the artifact itself. Custom Skills are instructions, never scripts — they can't execute arbitrary code.

Plugins

Capability bundles a model uses consistently.

Plugins package capabilities Axiom already has — and never grant one the current model or host doesn't actually expose.

Web Research Data Lab File Intelligence Connected Apps (MCP) Creator Studio

Themes, and a live map of what Axiom is doing.

Switch the whole app between themes instantly from Settings → General — no restart, remembered, and reaching every surface down to the native title bar. Neuron shows active sessions, tool usage and activity across Chat, Workplace, Documents, Study and Calculator.

Axiom's Settings, General tab, showing the theme picker. — **Axiom Dark** — the default theme, picked from Settings.

Axiom with the Gruvbox Dark theme applied. — **Gruvbox Dark** — applied instantly, everywhere.

Axiom's Neuron tab: a live map linking User to Chat, Workplace, Documents, Study and Calculator nodes. — **Neuron** — a live neural map of sessions and tool use.

Tools, memory and privacy.

Your data lives in %LOCALAPPDATA%\Axiom; debug runs use a separate profile. Nothing personal is ever packaged into a release.

Calculator	Scientific expressions and unit conversions
Python sandbox	Persistent session, bounded execution, chart capture
Java sandbox	Compile and run Java for supported Workplace tasks
Web search	Multi-source querying, deduplication, trust scoring, synthesis
Codebase access	Inspect and patch an explicitly connected workspace, with validation
Session memory	In-session episodic context for Workplace roles and study sessions
Persona memory	Persistent preferences stored locally
Context compaction	Keeps important requirements intact as conversations grow

With background operation and the system tray both on, closing the window hides Axiom in the tray instead of ending work in progress; Exit Axiom in the tray menu stops the process completely.

Get started in five steps.

Download the latest Windows ZIPFrom the Releases page — the button below always points at the newest one.
Extract the whole folderDon't run the executable from inside the ZIP.
Launch Malx_AI.exeSelf-contained; no .NET install needed.
Pick how it runs a modelImport or install a GGUF, configure Hybrid Local, or add an OpenRouter key.
Open Chat, or head to WorkplaceThat's it — updates arrive in-app after this.

Download for Windows

System requirements

OS: Windows 10 or 11 (64-bit)
RAM: 4 GB minimum; 16 GB recommended for local models
CPU: Modern x64
GPU: Optional NVIDIA CUDA acceleration
Runtime: Self-contained release

Updating

Axiom checks the stable Releases feed at startup. Use the in-app notice or Settings → General → Check for updates: it downloads and verifies the package, stages it outside the install directory, replaces only package-managed files after shutdown, and restarts in place.