AX

A fast, lightweight AI agent for the terminal, built in Rust.

One source workspace builds for Windows, Linux and macOS. See platform support for native build targets, shell settings, installation and capability limits; new platform archives require a release containing those assets before they can be downloaded.

English | 简体中文

AX is a small, native terminal agent. It starts fast, streams responses, and keeps sessions, memory, credentials and model config on your machine — no service to sign up for, no remote telemetry.

AX is the agent; AX Crew is the control plane built alongside it — it orchestrates and controls AX across multiple machines and keeps their work running when a machine goes offline. See ACP and Crew integration.

• Rust native — one binary for the terminal agent; no Node or Docker needed for core use. Optional Browser Use needs Node and Playwright.
• Independent UI permissions — app/site access and action approval are separate from the file/terminal sandbox. Computer Use defaults on and preserves explicit disable settings; Windows UIA, Linux AT-SPI and macOS Accessibility adapters have platform prerequisites and limits. See Computer Use and host access.
• Fast startup — providers, skills, MCP servers and memory load on demand.
• Local system resources — /system shows measured CPU/memory usage and supported NVIDIA GPU usage/VRAM; collection runs only when requested.
• Standard skills and focused tools — Agent Skills SKILL.md packages with lazy instructions and optional resources, read-only web search/fetch that batches concurrent queries and page fetches and deduplicates their results, LSP through configurable MCP servers, and native image input on supported vision models.
• Small binary — a compact workspace of focused crates, not a framework.
• Simple workflow — type, get an answer, switch models, move on.
• Local-first — sessions, memory and settings live in .ax beside the AX executable.

No benchmarks are published yet; "fast" here means the startup path does as little as possible by design.

By default, AX stores persistent data under <install-dir>/.ax, with separate project stores in projects/<project-key>/. Opening another working directory does not move data there; deleting a workspace does not delete its history. On first launch, legacy ~/.ax and known project .ax stores are copied without deleting the originals. AX_HOME and --data-dir remain explicit overrides.

Portable backups: ax export backup.axpack saves sessions and memory; use --memory or --sessions to select one type. Run ax import backup.axpack --dry-run to inspect conflicts, then ax import backup.axpack to merge. Credentials and caches are excluded. See backup details.

Quick Start

Installing and running AX

Mac or Linux:

curl -fsSL https://raw.githubusercontent.com/Axium-Labs/AX/main/scripts/install.sh | sh

Windows:

powershell -ExecutionPolicy Bypass -c "iex ((iwr 'https://raw.githubusercontent.com/Axium-Labs/AX/main/scripts/install.ps1' -UseBasicParsing).Content)"

Then simply run:

ax

Your first session:

/login      # sign in with a provider (Codex OAuth or an API key)
/model      # pick a model
hello       # just start typing

That's it. Your session is saved automatically, and the model you pick is remembered for next time.

Installation

The installer fetches the latest release from GitHub, verifies the archive against the release's SHA256SUMS, and installs a single ax / ax.exe binary:

• macOS / Linux — installs to ~/.local/bin by default and prints a PATH hint if that directory is not already on your PATH.
• Windows — installs to %LOCALAPPDATA%\Programs\AX\bin by default and adds it to your user PATH (only if it is not there already).

Both installers accept two environment variables:

To update the executable you are currently running to the latest GitHub Release, run ax --update. AX checks the release version, verifies the download against SHA256SUMS, and replaces that executable; it also refreshes the bundled payload from the same archive, installing any missing skill package into <install-dir>/.ax/skills and writing <install-dir>/.ax/mcp.toml only when no config exists yet (existing skills and configs are left alone). The bundled payload is refreshed on every ax --update, even when the binary is already up to date, so newly added skill packages are picked up without a version bump. On Windows, the verified update is scheduled for replacement after the command exits; restart AX before using the new version. The command does not otherwise change your <install-dir>/.ax user data or any project's .ax directory. If you installed a separate copy using AX_INSTALL_DIR, run that copy's ax --update to update it. A push to main alone is not a release; the release workflow runs for v* tags.

curl -fsSL https://raw.githubusercontent.com/Axium-Labs/AX/main/scripts/install.sh | sh
AX_INSTALL_DIR="$HOME/bin" curl -fsSL https://raw.githubusercontent.com/Axium-Labs/AX/main/scripts/install.sh | sh

powershell -ExecutionPolicy Bypass -c "iex ((iwr 'https://raw.githubusercontent.com/Axium-Labs/AX/main/scripts/install.ps1' -UseBasicParsing).Content)"

Uninstall

Removing AX is just deleting the binary (and, on Windows, the PATH entry the installer added). AX keeps its data separately in <install-dir>/.ax, which you can delete too if you want a clean slate.

macOS / Linux

rm -f ~/.local/bin/ax        # or wherever AX_INSTALL_DIR pointed
rm -rf ~/.local/bin/.ax                 # config, sessions, memory, model catalog (optional)

Windows

Remove-Item -Force "$env:LOCALAPPDATA\Programs\AX\bin\ax.exe"
# remove the installer's PATH entry (optional but tidy)
$p = [Environment]::GetEnvironmentVariable('Path', 'User')
[Environment]::SetEnvironmentVariable('Path', ($p -split ';' | Where-Object { $_ -notlike '*Programs\AX\bin*' }) -join ';', 'User')
Remove-Item -Recurse -Force "$env:LOCALAPPDATA\Programs\AX"   # deletes stored sessions and memory too (optional)

Development

Build from source

Requires Rust 1.92+.

git clone https://github.com/Axium-Labs/AX.git
cd AX
cargo build --release
# binary at target/release/ax.exe (Windows) or target/release/ax (Unix)

Providers & Models

AX discovers providers from what is already on your machine — credentials in <install-dir>/.ax/auth.json and standard environment variables like DEEPSEEK_API_KEY. No network call happens during startup.

Models are discovered dynamically into <install-dir>/.ax/models/ (a bundled catalog plus a per-provider refresh cache). Switch anytime with /model; the last selection — provider, model and reasoning effort — is persisted to <install-dir>/.ax/config.json and restored on the next launch.

To pick a provider explicitly from the command line:

ax --provider deepseek
ax --provider openai-codex --model <model-id>
ax run "explain this repo" --provider deepseek

--provider takes a provider id from the catalog, not only the deepseek / openai / codex / compatible aliases. Name the id whenever a vendor is listed more than once — Xiaomi's plain entry and its CN/SGP/AMS token-plan regions share model ids but not endpoints:

ax run "explain this repo" --provider xiaomi-token-plan-cn --model mimo-v2.5

Usage

Start the UI (default):

ax

Run one task and exit (non-interactive):

ax run "summarize the TODOs in ./src"

Run independent tasks with bounded concurrency:

ax agents "review pr #12" "write changelog" --concurrency 2

To expose AX's existing runtime to an ACP client, run ax acp. To connect this machine to an AX Crew backend, run ax crew pair <code> --gateway <https-url> once, then ax crew connect <https-url>. See ACP and Crew integration.

Design

Fast. Lightweight. Simple.

AX deliberately does not load everything up front. Skills, MCP servers, providers and memory are all activated only when the task actually needs them — so the things that make an agent powerful never get in the way of starting it or of a normal conversation. The trade-off is intentional: capability is loaded on demand, not at startup.

Documentation

The internal architecture — agent loop, context management, multi-agent scheduling, SQLite schema, module layout — is documented in docs/README.md. Start with docs/architecture.md.

Contributing

See CONTRIBUTING.md for reporting issues, developing changes, submitting PRs and the review process.

Acknowledgments

• [AtomGit / GitCode](https:/