fak

fak fixes mistyped shell commands with an AI model, shows the proposed change, and only executes it after you approve it.

The release binaries contain the LFM2.5 230M GGUF model. Local mode therefore works without Ollama, an external model server, or a separate model download at runtime.

Features

• Embedded local LFM2.5 model for offline command correction.
• Optional NVIDIA NIM, Cerebras, Gemini, and Groq providers.
• A visible command diff and confirmation prompt before execution.
• Bash, Zsh, Fish, Nushell, and PowerShell history integration.
• Release binaries for Linux, macOS, and Windows on x86_64 and ARM64.
• Rust code compiled with unsafe forbidden.

Download a release

Download the standalone binary or the archive for your operating system and CPU from the latest release.

The standalone files are the actual compiled binaries and are published directly as release assets:

Platform	Standalone binary	Archive
Linux x86_64	fak-linux-x86_64	fak-linux-x86_64.tar.gz
Linux ARM64	fak-linux-aarch64	fak-linux-aarch64.tar.gz
macOS Intel	fak-macos-x86_64	fak-macos-x86_64.tar.gz
macOS Apple Silicon	fak-macos-aarch64	fak-macos-aarch64.tar.gz
Windows x86_64	fak-windows-x86_64.exe	fak-windows-x86_64.zip
Windows ARM64	fak-windows-aarch64.exe	fak-windows-aarch64.zip

Every standalone binary and archive has a matching .sha256 file. Archives also contain LICENSE-MODEL.txt.

Install with one command

The installers download the latest release for your operating system and CPU, verify its SHA-256 checksum, install it for your user, add it to PATH, and enable the shell history hook. They do not require Rust, Ollama, or a separate model download.

On Linux or macOS, run this from Bash, Zsh, Fish, or Nushell:

curl --proto '=https' --tlsv1.2 -sSf https://raw.githubusercontent.com/Ferrb9579/fak/master/install.sh | sh

Open a new terminal after the installer finishes. It configures the startup file for Bash, Zsh, Fish, or Nushell as appropriate, so this works immediately in the new terminal:

git statuss
fak

On Windows, run this in PowerShell:

irm https://raw.githubusercontent.com/Ferrb9579/fak/master/install.ps1 | iex

The Windows installer adds fak.exe to your user PATH and adds a native fak function to your PowerShell profile. Git Bash and WSL users should run the Unix installer from Git Bash or WSL instead.

The shell integrations are:

Shell	Automatic setup	Startup file
Bash	Yes	~/.bashrc, ~/.bash_profile, or ~/.profile
Zsh	Yes	~/.zshrc
Fish	Yes	~/.config/fish/config.fish
Nushell	Yes	~/.config/nushell/config.nu and fak.nu
PowerShell	Yes	$PROFILE

The release binaries themselves run on Linux, macOS, and Windows for x86_64 and ARM64. Shell integration is separate from the operating system: use the installer for the shell you actually open.

The installers use the latest published release by default. To install a specific release, set FAK_VERSION before running the installer:

FAK_VERSION=v1.0.2 sh -c 'curl --proto "=https" --tlsv1.2 -sSf https://raw.githubusercontent.com/Ferrb9579/fak/master/install.sh | sh'

If you download the Windows archive manually, the executable is inside the ZIP. After extraction, run:

fak-windows-x86_64\fak.exe --help

The ARM64 archive contains the same fak.exe name under fak-windows-aarch64.

Manual install on Linux or macOS

For example, the following installs the standalone Linux x86_64 binary into ~/.local/bin:

curl -LO https://github.com/Ferrb9579/fak/releases/latest/download/fak-linux-x86_64
mkdir -p ~/.local/bin
install -m 0755 fak-linux-x86_64 ~/.local/bin/fak

Replace fak-linux-x86_64 with the matching macOS or ARM64 standalone asset from the table above. The archive form is also available when you want the model license beside the executable:

curl -LO https://github.com/Ferrb9579/fak/releases/latest/download/fak-linux-x86_64.tar.gz
tar -xzf fak-linux-x86_64.tar.gz
mkdir -p ~/.local/bin
install -m 0755 fak-linux-x86_64/fak ~/.local/bin/fak

Make sure ~/.local/bin is on your PATH:

export PATH="$HOME/.local/bin:$PATH"
fak --help

Manual install on Windows

In PowerShell, you can download the standalone .exe directly:

Invoke-WebRequest `
  -Uri https://github.com/Ferrb9579/fak/releases/latest/download/fak-windows-x86_64.exe `
  -OutFile fak.exe
.\fak.exe --help

Alternatively, download the archive from the release page and extract it:

Expand-Archive .\fak-windows-x86_64.zip -DestinationPath .
.\fak-windows-x86_64\fak.exe --help

You can move the extracted folder to a permanent location and add that folder to your user PATH if you want to run fak.exe from anywhere. For automatic history integration, use the one-command installer above. It adds a native PowerShell hook; Git Bash and WSL use the Bash/Zsh hook.

Verify a download

For a standalone Linux binary:

sha256sum -c fak-linux-x86_64.sha256

For a standalone macOS binary:

shasum -a 256 -c fak-macos-aarch64.sha256

On Windows, compare the output of Get-FileHash with the contents of the matching standalone checksum file:

(Get-FileHash .\fak-windows-x86_64.exe -Algorithm SHA256).Hash
Get-Content .\fak-windows-x86_64.exe.sha256

The archive checksum files use the same commands with the archive filename instead.

Choose a model provider

Use the embedded model when you do not want an API key or external service:

export FAK_PROVIDER=local
fak git stats

The local provider is bundled in the release binary. It does not require Ollama or a model installation.

Remote providers can be selected explicitly with FAK_PROVIDER:

Provider	Selection	API key	Optional model variable
Local	local or lfm	None	—
NVIDIA	nvidia	NVIDIA_API_KEY	NVIDIA_MODEL
Cerebras	cerebras	CEREBRAS_API_KEY	CEREBRAS_MODEL
Gemini	gemini	GEMINI_API_KEY	GEMINI_MODEL
Groq	groq	GROQ_API_KEY	GROQ_MODEL

For example, NVIDIA configuration:

export FAK_PROVIDER=nvidia
export NVIDIA_API_KEY="your-nvidia-api-key"
fak git stats

Default remote models are:

• NVIDIA: google/gemma-4-31b-it
• Cerebras: llama3.1-8b
• Gemini: gemma-4-31b-it
• Groq: llama-3.3-70b-versatile

If FAK_PROVIDER is not set, fak uses the embedded local model. Remote providers are never selected automatically; set FAK_PROVIDER explicitly when you want one.

Never commit API keys. For local development, fak can read an ignored .env file, but normal shell environment variables work as well:

FAK_PROVIDER=local
# FAK_PROVIDER=nvidia
# NVIDIA_API_KEY=replace-me

Automatic shell integration

fak --alias prints a shell function that connects fak to your command history. The one-command installer adds it automatically. If you install a binary manually, add the evaluation command to your shell startup file.

For Bash, add it to ~/.bashrc:

echo 'eval "$(command fak --alias)"' >> ~/.bashrc
source ~/.bashrc

For Zsh, add it to ~/.zshrc:

echo 'eval "$(command fak --alias)"' >> ~/.zshrc
source ~/.zshrc

Run the append command only once. If the line is already present, just open a new terminal or run source ~/.bashrc / source ~/.zshrc.

If your Bash terminal is configured as a login shell, place the same line in ~/.bash_profile or ~/.profile instead. Git Bash uses the Bash instructions.

For Fish, add this to ~/.config/fish/config.fish:

command fak --alias | source

For Nushell, generate a hook file and source it from ~/.config/nushell/config.nu:

fak --alias | save --force ~/.config/nushell/fak.nu
source ~/.config/nushell/fak.nu

For PowerShell, the Windows installer adds the hook automatically. For a manual setup on Windows, run:

Invoke-Expression (& fak.exe --alias)

The generated hook reads the current shell's history, asks the embedded model for a correction, and executes it in that same shell. Fish replaces the interactive command line before executing the correction, so the corrected command is retained by Fish history.

To enable it only in the current terminal without editing a startup file:

eval "$(command fak --alias)"

Uninstall

On Linux or macOS, remove only the installed binary:

rm -f ~/.local/bin/fak

Then remove the exact hook line from any startup file where the installer added it, such as ~/.bashrc, ~/.bash_profile, ~/.profile, ~/.zshrc, ~/.config/fish/config.fish, or ~/.config/nushell/config.nu. Remove ~/.config/nushell/fak.nu if it exists. Remove the PATH line too if it was added only for fak. Open a new terminal afterward.

On Windows, the one-command installer uses $env:LOCALAPPDATA\Programs\fak. Remove it and remove that directory from your user PATH:

Remove-Item "$env:LOCALAPPDATA\Programs\fak" -Recurse -Force

Also remove the block between # >>> fak shell integration >>> and # <<< fak shell integration <<< from $PROFILE. For Git Bash or WSL, remove the generated hook line from the relevant Bash or Zsh startup file.

Use it from a supported shell

Now correct the most recent command:

gti status
fak

Or provide a command directly:

fak git stats

fak will display the proposed correction. Press Enter to execute it, or press Ctrl+C to cancel. The accepted command is added to shell history.

For commands containing pipes, redirects, or other shell syntax, quote the complete command:

fak 'git status | head'

You can generate an alias with another function name if fak is already in use:

echo 'eval "$(command fak --alias=fix)"' >> ~/.bashrc
source ~/.bashrc
fix git stats

The same custom-alias command can be placed in ~/.zshrc. For Fish, use command fak --alias=fix | source in ~/.config/fish/config.fish. For Nushell, generate the file with fak --alias=fix | save --force ~/.config/nushell/fak.nu and source it from config.nu. Replace fix with the function name you want.

Safety and privacy

• Review the proposed command before pressing Enter. The generated text is executable shell input.
• Local mode keeps the request on your machine and does not need a network service.
• Remote providers receive the command and prompt through their APIs; check their privacy policies before using them with sensitive commands.
• API keys are read from environment variables and are not embedded in the binary.
• Each release archive includes LICENSE-MODEL.txt for the bundled model. Review that license before redistributing the binary.

Build from source

Release archives are the easiest way to use fak. A source build requires Rust stable plus the native C/C++ build tools used by llama-cpp-2 on your platform.

The model file must exist at this exact path before compiling:

models/LFM2.5-230M.Q4_K_M.gguf

Download the pinned LFM2.5-230M-Q4_K_M.gguf file from the LFM2.5 GGUF repository, rename it to the path above, and verify it against the SHA-256 value in the release workflow before building:

cargo build --release

The resulting binary is target/release/fak on Unix-like systems and target/release/fak.exe on Windows.

Troubleshooting

FAK_HISTORY is missing

The shell function has not been loaded. Use the command for your shell:

eval "$(fak --alias)"

command fak --alias | source

fak --alias | save --force ~/.config/nushell/fak.nu
source ~/.config/nushell/fak.nu

Invoke-Expression (& fak.exe --alias)

Then run fak again.

An API key error appears

An API key is only needed when a remote provider was selected explicitly. Return to the embedded model with:

unset FAK_PROVIDER
export FAK_PROVIDER=local

The model is missing during a source build

Confirm that the downloaded GGUF file is named exactly:

models/LFM2.5-230M.Q4_K_M.gguf