Install
Every program is one self-contained binary. There is nothing else to install: no Python, no C library, no CUDA toolkit. GPU backends are loaded at run time from the driver already on the machine.
Download
Section titled “Download”The download for this machine comes first; every other platform is listed after it.
For your machine
macOS
jitllm for Apple silicon (.dmg)
- Intel Mac (.dmg)
- The app alone, zipped: Apple silicon, Intel
- Homebrew:
brew tap jitllm/jitllm && brew trust jitllm/jitllm && brew install --cask jitllm
Windows
Linux
curl -fsSL https://jitllm.org/install.sh | sh
Other platforms
Every archive, with checksums.txt: the latest release.
macOS: the app
Section titled “macOS: the app”Download jitllm_<version>_darwin_arm64.dmg for Apple silicon, or jitllm_<version>_darwin_amd64.dmg for Intel, from the latest release. Open it and drag jitllm onto Applications.
The same signed app also comes zipped, without the disk image: jitllm_<version>_darwin_arm64.app.zip or _amd64.app.zip. Double-click it to unzip and move jitllm to Applications.
The app carries the CLI and the server inside it. On first launch it offers to link jitllm and jitllmd into /usr/local/bin (asking for your password if that folder needs it, or using ~/.local/bin if you decline). Settings > Command line links or unlinks them later, and Start the server at login runs jitllmd in the background on localhost:4242 through a launchd agent, until you turn it off.
With Homebrew:
brew tap jitllm/jitllmbrew trust jitllm/jitllmbrew install --cask jitllm # the app, with jitllm and jitllmd on PATHbrew install jitllm # the CLI and the server onlyThe formula’s server also runs as a Homebrew service: brew services start jitllm.
Windows: the setup
Section titled “Windows: the setup”Run jitllm-setup_<version>_windows_amd64.exe (or _arm64.exe on Arm) from the latest release. It installs for your account only and never asks for an administrator:
jitllm,jitllmdand the desktop app in%LOCALAPPDATA%\Programs\jitllm, which is added to your userPATH;- jitllm in the Start menu;
- optionally, Start the server at login:
jitllmdonlocalhost:4242, with no console window; - an entry under Settings > Apps whose uninstaller removes all of it. Your models and chats are kept.
The portable desktop app
Section titled “The portable desktop app”jitllm-desktop_<version>_windows_amd64.exe (or _arm64.exe) is the desktop app alone, one file that runs from wherever you save it and installs nothing: no Start menu entry, no PATH, no uninstaller. It keeps its settings and chats in the same folder as an installed copy (%APPDATA%\jitllm), so the two share them. Delete the file to remove it.
First launch of an unsigned build
Section titled “First launch of an unsigned build”Until a release is signed with a Developer ID (macOS) and a code-signing certificate (Windows), each system warns once:
- macOS says “jitllm” cannot be opened because Apple cannot check it for malicious software, or that it is damaged. Control-click the app in Applications and choose Open, then Open again; or in System Settings > Privacy & Security choose Open Anyway. A Homebrew install can drop the quarantine with
xattr -dr com.apple.quarantine /Applications/jitllm.app. - Windows SmartScreen says Windows protected your PC. Choose More info, then Run anyway.
A signed and notarized release opens with neither warning.
From a script
Section titled “From a script”On Linux and macOS:
curl -fsSL https://jitllm.org/install.sh | shOn Windows, in PowerShell:
irm https://jitllm.org/install.ps1 | iexThe script finds the latest release, downloads the archives for this machine, checks each against the release’s checksums.txt, and installs the binaries:
| Linux and macOS | Windows | |
|---|---|---|
| where | ~/.local/bin, or /usr/local/bin as root |
%LOCALAPPDATA%\Programs\jitllm, added to your user PATH |
| which programs | arguments: sh -s -- all |
$env:JITLLM_PROGRAMS = "all" before the line |
| a version | JITLLM_VERSION=v0.1.0 |
$env:JITLLM_VERSION = "v0.1.0" |
| another directory | JITLLM_INSTALL_DIR=... |
$env:JITLLM_INSTALL_DIR = "..." |
| where the macOS app goes | JITLLM_APPS_DIR=... (default /Applications, or ~/Applications when that is not writable) |
On Linux the script also installs jitllmd as a systemd user service on localhost:4242 (jitllmd service install), so jitllm run NAME and any OpenAI client find a server already running; sh -s -- --no-service (or JITLLM_NO_SERVICE=1) skips it. Its log is journalctl --user -u jitllm; jitllmd service status|stop|start|uninstall manage it, on every system.
The programs are jitllm (the CLI) and jitllmd (the server), installed by default, and desktop (the desktop app) and tui (the terminal app); all is the four. On macOS, desktop installs the signed jitllm.app from the release’s .app.zip rather than a bare binary. For example, the CLI and the terminal app:
curl -fsSL https://jitllm.org/install.sh | sh -s -- jitllm tuiRun it again to update. To install by hand, the latest release has one archive per program and platform, named <program>_<version>_<os>_<arch>: jitllm, jitllmd, jitllm-desktop and jitllm-tui, for linux, darwin and windows on amd64 and arm64 (.zip on Windows, .tar.gz elsewhere), with checksums.txt beside them. Unpack one and put the binary on your PATH. The server also ships as a container image; see Docker.
A GPU needs only its driver:
| Backend | What jitllm loads | Where it comes from |
|---|---|---|
| CUDA | libcuda.so.1 (Linux, WSL), nvcuda.dll (Windows) |
the NVIDIA driver |
| Vulkan | libvulkan.so.1, vulkan-1.dll, or libvulkan.1.dylib / MoltenVK |
the GPU driver or Vulkan loader |
| Metal | the system Metal framework | macOS on Apple silicon |
jitllm generates its own PTX, SPIR-V and Metal shaders, so the CUDA toolkit, nvcc and a shader compiler are never needed. With no usable GPU, automatic mode runs on the CPU.
Build from source
Section titled “Build from source”With Go 1.26 or newer and Git, and no C toolchain (there is no cgo):
git clone https://github.com/jitllm/jitllm.gitcd jitllm
# The engine and its command line.go install ./cmd/jitllm
# The API server. It is a separate module, so the core keeps one dependency.(cd server && go install ./cmd/jitllmd)
# The desktop app and the terminal app, modules of their own too.(cd ui && CGO_ENABLED=0 go build -o "$(go env GOPATH)/bin/jitllm-desktop" .)(cd tui && go build -o "$(go env GOPATH)/bin/jitllm-tui" .)These put the binaries in Go’s bin directory, $(go env GOPATH)/bin unless GOBIN says otherwise. Make sure it is on your PATH:
export PATH="$PATH:$(go env GOPATH)/bin"On Windows the same commands produce .exe files; in PowerShell run the second as cd server; go install .\cmd\jitllmd; cd ...
The core module has a single dependency, goffi, which is how it calls the GPU drivers without cgo. Builds are static and cross-compile like any Go program:
GOOS=windows GOARCH=arm64 go build ./cmd/jitllmGOOS=darwin GOARCH=arm64 go build ./cmd/jitllmCheck what jitllm sees
Section titled “Check what jitllm sees”jitllm hardwarehardware prints the CPU’s instruction-set tier and the kernels it will generate, every GPU each backend can open, and how much memory is actually spendable on each. One card that answers to two backends (CUDA and Vulkan, say) is listed under both and counted once. Use the device names it prints with -devices.
jitllm doctor answers the shorter question, whether this machine and build can run at all, and given a container (jitllm doctor model.jlm) whether it can run that one. It names each thing missing with what to do next, and exits 0 when all is well and 2 when something is missing; -json prints the report for a script or an agent.
The apps
Section titled “The apps”The desktop app and the terminal app run the engine in their own process and share their settings and saved chats. Install them with the script’s desktop and tui, or build them as above.
- Quickstart: fetch a model and serve it.
- Hardware support: what each CPU tier and GPU backend runs.