omm — Open source Model Manager

omm is an apt/brew-style package manager for local LLMs (GGUF). It installs models into a central hub, links them into seven local AI runners automatically (Ollama, LM Studio, Jan, AnythingLLM, Msty, text-generation-webui, KoboldCpp), and can recommend a model that fits your hardware.

Install

Verified Git-source installer (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/omm-hippo/omm/main/install.sh | sh

This bootstraps python3, git, and pipx if missing (Debian/Ubuntu via apt, or Homebrew on macOS), then installs omm as an isolated CLI via pipx. Open a new shell afterward so your PATH picks up omm.

PyPI (macOS, Linux, and Windows):

python -m pip install omm-model

This does not go through the signed-commit verification described below; it relies on PyPI's own account security and TLS, the same trust model as installing any other PyPI package.

For an isolated command-line installation, pipx is recommended:

pipx install omm-model

The distribution name is omm-model; the installed command and Python import remain omm. Upgrade and remove it with the same tool that installed it:

python -m pip install --upgrade omm-model
python -m pip uninstall omm-model

# Or, for pipx:
pipx upgrade omm-model
pipx uninstall omm-model

Homebrew Tap (macOS):

brew install omm-hippo/omm/omm

Upgrade or remove the formula with Homebrew. Removing the formula preserves downloaded models and settings under OMM_HOME:

brew upgrade omm-hippo/omm/omm
brew uninstall omm-hippo/omm/omm

The Homebrew formula and PyPI package can move on separate release schedules; use brew info omm-hippo/omm/omm to see the version currently provided by the Tap. omm update does not modify a Homebrew installation and instead prints the matching brew upgrade command.

Verified Git-source installer (Windows PowerShell):

# This must run before irm: script-internal TLS settings are too late for its first download.
[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12; irm https://raw.githubusercontent.com/omm-hippo/omm/main/install.ps1 | iex

This bootstraps Python and git via winget if missing (built into Windows 10 2004+ and Windows 11 — on older Windows, install Python 3.10+ and git manually first), then installs omm through that exact validated Python interpreter. Open a new PowerShell window afterward so your PATH picks up omm. On Windows, model exposure tries an unprivileged same-volume hard link first, then a symbolic link (Developer Mode or Administrator), then an owned copy. Before copying, omm checks destination free space and reports that the model now consumes additional bytes. File junctions do not apply because model targets are files, not directories.

Requirements: Python 3.10+. The optional NVIDIA detector is installed only when nvidia-smi indicates an NVIDIA driver.

Supported platforms

omm is tested in CI on Windows, macOS, and Linux with Python 3.10+. Windows 10 22H2/11 is the supported Windows baseline because that matches Ollama's native Windows requirements. Hardware scan, install, linking, benchmark, update, and contribution flows are cross-platform; Ollama remains the only benchmark engine.

Both installers clone to a versioned staging directory, verify the signed commit against a bootstrap trust anchor, and only then switch pipx to it. Do not replace this with an unverified git clone plus pipx install if commit authenticity matters.

omm update updates only a canonical OMM Git-source installation. For a PyPI or pipx installation it leaves files unchanged and prints the matching package manager upgrade command. The Git-only beta channel is likewise unavailable to package-managed installations.

Package-channel verification

Installation path	Highest verified level	Remaining limitation
PyPI / pipx	Simulator-verified on GitHub-hosted Windows, macOS, and Ubuntu runners using the public package	A real upgrade from the first public release remains a separate user-path check
Homebrew Tap	Physical-device-verified on an Apple Silicon Mac for public Tap install, omm --version, brew test, upgrade guidance, and uninstall	Intel Mac installation is not yet physical-device-verified

Additional package-manager commands are added here only after their public registry path has been installed and verified.

Local AI runners

The first bare omm run on a fresh install (or omm setup any time after) shows a hardware summary and a checklist of local AI runners. Checking one that omm knows how to install runs its official installer with live progress in the terminal; checking one it doesn't yet automate on your platform prints a link instead. Automation coverage today:

Runner	Automated on	Manual elsewhere
Ollama	macOS, Linux, Windows	—
LM Studio	macOS, Linux, Windows (headless lms CLI)	—
Jan	macOS (Homebrew), Windows (winget), Linux (Flatpak)	wherever that package manager isn't installed
AnythingLLM	macOS (Homebrew), Windows (winget)	Linux
Msty	macOS (Homebrew)	Windows, Linux
KoboldCpp	macOS (Apple Silicon), Linux (x86_64), Windows	Intel Mac, other architectures
text-generation-webui	macOS (any arch), Linux/Windows (x86_64)	ARM Linux/Windows

Every currently-installed runner is also listed (marked as already installed, not selectable) rather than hidden, so the checklist always reflects what omm actually detects on the machine.

Storage location

The model hub and omm state default to ~/.omm. Set OMM_HOME before installation and on later runs to put them on another volume:

[Environment]::SetEnvironmentVariable("OMM_HOME", "D:\omm", "User")
$env:OMM_HOME = "D:\omm"

export OMM_HOME=/mnt/models/omm

Ollama's own model location follows OLLAMA_MODELS. LM Studio follows its home pointer; set OMM_LMSTUDIO_MODELS_DIR when LM Studio uses a custom directory that omm cannot discover automatically.

Completion and uninstall

Install native shell completion once, then restart the shell:

omm --install-completion powershell

omm --install-completion bash  # or zsh/fish

For a PyPI installation, use python -m pip uninstall omm-model; for pipx, use pipx uninstall omm-model. Both commands preserve downloaded models and settings under OMM_HOME.

To remove a Git-source installer installation while preserving downloaded models and settings:

irm https://raw.githubusercontent.com/omm-hippo/omm/main/uninstall.ps1 | iex

curl -fsSL https://raw.githubusercontent.com/omm-hippo/omm/main/uninstall.sh | sh

Run a downloaded script with -Purge (PowerShell) or --purge (sh) to remove the model hub and settings too. Purge removes only known omm-owned paths and leaves unrelated files in a custom OMM_HOME untouched. Installers mark custom homes so uninstallers can refuse ambiguous or unsafe locations; shell profiles are never rewritten during uninstall.

Usage

omm setup  # First-run setup wizard: hardware scan + engine checklist (re-runnable any time)
omm scan [--json]  # Print a hardware, runner, and model summary (RAM, VRAM, OS)
omm recommend [--json]  # Suggest a model that fits this machine, then offer to install it
omm tune <name> [--json]  # Recommend context, GPU offload, threads, and batch size
omm benchmark <name>...  # Local quality + speed smoke evidence for one or more installed models
omm search <query> [--json] [--skip-unfit] [--limit N] [--provider curated|huggingface|modelscope]  # Search curated models, cached candidates, and HuggingFace
omm install <name> [--skip-unfit] [--upload/--no-upload] [--force] [--verify-runtime|--no-verify-runtime]  # Download, link, and optionally verify a model
omm import [directory] [--yes]  # Adopt GGUF files already sitting in Ollama/LM Studio (or a given directory) into the hub
omm uninstall <name> [--dry-run]  # Uninstall a model and clean up its symlinks/manifests (alias: rm)
omm uninstall all [--yes] [--dry-run]  # Uninstall every model installed via omm
omm list [--json] [--engine NAME]  # Show models installed via omm and their linked status (alias: ls)
omm info <name> [--json]  # Show a model's name, version, size, and linked-program run commands
omm verify <name> [--engine ollama|lmstudio] [--keep-loaded]  # Prove local load + generation works
omm upgrade <name> [--dry-run]  # Refresh a model against its source if it has changed since install (alias: up)
omm upgrade [--yes] [--dry-run]  # Check every installed model for updates
omm link [--engine NAME]  # Re-verify and repair every installed model's LM Studio/Ollama links
omm link <directory>  # Reuse central GGUF files; Windows warns if a real copy is required
omm autoremove  # Clean up broken symlinks and orphaned partial downloads
omm contribute [--yes]  # Repeatedly install/benchmark/upload hardware-fit models to grow the dataset
omm update  # Update a canonical OMM Git-source install; package installs print their manager command
omm setting  # Interactive menu for telemetry, upload policy, error reports, version, theme, calibration, and catalog trust
omm setting version [--stable|--beta]  # Show or switch the update channel `omm update` pulls from
omm setting telemetry --endpoint <url>  # Configure where benchmark telemetry is sent
omm setting upload --enable|--disable|--ask  # Configure the benchmark-upload send policy
omm setting error-reports --enable|--disable|--ask  # Configure the opt-in crash/error-report send policy
omm setting memory-guard --policy ask|block|observe  # Protect Ollama loads from live memory pressure
omm setting theme [--set NAME]  # Show or change omm's output color theme
omm setting calibrate <name>  # Locally correct predicted speed with an installed Ollama model
omm setting catalog-trust --manifest-url <url> --public-key <key>  # Require signed recommendation downloads
omm setting catalog-status  # Show signed recommendation data and rollback snapshots
omm setting catalog-rollback  # Restore the most recent different recommendation snapshot
omm help [command]  # Show help, same as --help

install, uninstall, info, and upgrade accept either a model name/reference or the numeric index shown by the last omm search or omm list run in that terminal. search/install mark models predicted not to run on this machine's hardware in red.

omm verify checks more than a link: it asks before loading an unloaded model, sends one short deterministic prompt to a server already running on this computer, requires a non-empty answer, and releases only a model that OMM loaded for the check. It never starts Ollama or LM Studio, deletes the model, or stores the generated answer. LM Studio API authentication reads LM_API_TOKEN from the process environment and never writes it to config.json. Compatibility status is stored locally in models.json and is shown by omm info.

Scripting

All errors, warnings, and confirmation prompts print to stderr; --json output (supported on search/list/info/benchmark/tune/scan/recommend) is the only thing written to stdout, so it's safe to pipe (e.g. omm list --json | jq .). Any command that would otherwise prompt for confirmation fails fast with a non-zero exit code when there's no terminal attached instead of hanging — pass --yes/-y (works on every command that has a confirmation prompt) or the relevant flag (install --skip-unfit, install --upload/--no-upload) to run it unattended.

Four global flags work either before or after the subcommand name (omm --json search foo and omm search foo --json are equivalent): --json (structured output, where supported — see above), --yes/-y (skip confirmation prompts), --quiet/-q (suppresses progress bars and background status/hint lines — e.g. download progress, "Verifying checksum...", scan's "Run: omm link" nudge; errors, warnings, and the result of what you asked for still print), and --no-color (disable ANSI colors on omm's own console output and its download progress bar; the NO_COLOR environment variable does the same). Passing --json or --yes to a command that doesn't use them prints a warning to stderr instead of silently doing nothing. Exit codes are consistent across every command: 0 success, 1 failure, 2 usage error (bad flag/argument).

rm, ls, and up are short aliases for uninstall, list, and upgrade.

Set OMM_HOME to store everything (models, config, catalog history) under a different directory instead of ~/.omm — useful when $HOME's filesystem doesn't have room for GGUF models, e.g. OMM_HOME=/mnt/data/omm omm contribute --yes.

omm contribute refuses to start unless every model volume has at least 10 GiB free. Before each download it also budgets the central GGUF, a worst-case full Ollama import copy, any required Windows cross-volume copies, and safety headroom. Each model evaluation prints a heartbeat every 30 seconds and is terminated after an absolute 10-minute deadline instead of hanging an unattended session indefinitely.

Localfit does not assume all installed memory belongs to the model. A live scan subtracts memory currently used by other applications, keeps at least 2 GB (or 10% of RAM) for the OS and newly opened apps, and applies total-memory caps. Recommendation fit and omm tune use this safe budget, so rerunning a command adapts after memory-heavy applications are opened or closed.

omm benchmark runs a versioned eight-item bilingual arithmetic smoke pack against models already installed in Ollama. It stores parsed answers, correctness, pinned model metadata, and fixed-length timings under ~/.omm/evaluations/; it stores no generated text. Opt-in telemetry sends a locally computed CPU chip score (and GPU chip score, when a GPU is present) plus architecture and core counts — never the raw CPU/GPU model name — so speed predictions can distinguish otherwise identical Linux x86_64 machines. Results are uploaded only after explicit opt-in. The pack is intentionally small and is not a leaderboard.

On Windows, Ollama is detected by its HTTP API first, so a freshly installed tray app works even before the current terminal receives the new PATH. When the daemon is stopped, omm also checks Ollama's documented %LOCALAPPDATA%\Programs\Ollama location. It only stops daemon processes it started itself. Before deleting a contribution model, omm requests an Ollama unload, waits for /api/ps to confirm handle release, and uses bounded retries for Windows file locks. Real-time antivirus can still delay a first load; the benchmark uses repeated samples and reports their median. Do not disable your antivirus for omm.

Self-hosted benchmark data

Benchmark results are never uploaded without explicit per-run consent or an always policy. New installations include the hosted Firebase endpoint as the default destination, while existing local-only configurations stay local. To run the bundled FastAPI + SQLite collector instead:

pip install -e ".[server]"
export LOCALFIT_DB_PATH="$PWD/localfit.db"
export LOCALFIT_ADMIN_TOKEN="replace-with-a-long-random-token"
localfit-server

Explicitly configure the endpoint and opt in before uploading:

omm setting telemetry --endpoint http://127.0.0.1:8000/v1/benchmarks
omm setting upload --enable

Training can consume the authenticated export directly:

export LOCALFIT_ADMIN_TOKEN="replace-with-a-long-random-token"
python scripts/train_model.py \
  --telemetry-url http://127.0.0.1:8000/v1/benchmarks/export

Firebase Realtime Database JSON endpoints remain supported. An official *.firebaseio.com or *.firebasedatabase.app .json URL can be read without an admin token; self-hosted raw export requires LOCALFIT_ADMIN_TOKEN. Exact duplicate events are ignored.

Automated retraining is fail-closed. Configure LOCALFIT_TELEMETRY_EXPORT_URL; configure LOCALFIT_ADMIN_TOKEN as well for a self-hosted export (it is optional for an official Firebase JSON URL). The scheduled job otherwise stops without changing the published artifact. It requires at least 100 distinct valid v6/v7 configurations with explicit runtime and CPU metadata (legacy rows do not satisfy this minimum), rejects datasets with more than 25% invalid rows, and reserves a deterministic 20% holdout. A 64-tree v4 candidate replaces the incumbent only when both holdout RMSLE and P90 absolute percentage error stay within the configured regression limits. Selection is evaluated on whole hardware/request contexts, so sibling model variants never leak across training and holdout sets. Publishing also requires at least three multi-model selection groups plus complete top-1, regret, balanced-fit, and false-positive evidence. Missing evidence fails the gate. The artifact records the complete candidate/baseline evaluation report.

The same gate can validate an exported local dataset without contacting the collector:

python scripts/train_model.py --offline \
  --telemetry-file benchmarks.jsonl \
  --quality-gate --minimum-real-configurations 100 \
  --baseline published/localfit-recommend-model.json \
  --output candidate.json --quality-report quality-report.json

Synthetic bootstrap training remains available for local development, but the scheduled publishing workflow never uses it as a substitute for missing real benchmark data.

Signed recommendation data

omm setting catalog-trust --manifest-url <https-url> --public-key <base64-key> enables Ed25519 verification for future recommendation downloads. Existing artifacts are snapshotted before replacement and omm setting catalog-rollback restores the most recent different snapshot.

Development

pip install -e ".[dev]"
pytest

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup, testing, and PR conventions, and CODE_OF_CONDUCT.md for community expectations. Report security issues per SECURITY.md rather than as a public issue.

License

MIT — see LICENSE. Third-party dependency licenses are listed in THIRD_PARTY_NOTICES.md.