All plugins

omlx

Managed oMLX runtime and MLX model library for Hermes Desktop on Apple Silicon. Disclosure — on explicit setup downloads the oMLX 0.7.2 macOS app (SHA-256-pinned) from the official jundot/omlx GitHub release and runs it as a long-running loopback server (127.0.0.1, bearer key) under HERMES_HOME/plugin-data/omlx; downloads public Hugging Face model snapshots the user names; generates a local API key stored as HERMES_OMLX_LOCAL_API_KEY in the profile's .env; "Use" rewrites the profile's default model in config.yaml.

vitaliy-blprntModelsv0.1.0Updated 1 hour ago

About

Managed oMLX runtime and MLX model library for Hermes Desktop on Apple Silicon. Disclosure — on explicit setup downloads the oMLX 0.7.2 macOS app (SHA-256-pinned) from the official jundot/omlx GitHub release and runs it as a long-running loopback server (127.0.0.1, bearer key) under HERMES_HOME/plugin-data/omlx; downloads public Hugging Face model snapshots the user names; generates a local API key stored as HERMES_OMLX_LOCAL_API_KEY in the profile's .env; "Use" rewrites the profile's default model in config.yaml.

oMLX for Hermes

Managed local MLX models in Hermes Desktop, without a Hermes fork or core patches.
Requires Hermes 0.23.0+, macOS 15+, and Apple Silicon on the machine running the gateway.

What it includes

  • Install the checksum-pinned, self-contained oMLX 0.7.2 runtime.
  • Turn the private, authenticated loopback server on and off.
  • Download public Hugging Face model repositories, with byte progress and cancellation.
  • Browse model sizes, validate real generation before choosing a default, and confirm deletions.
  • Use the named omlx provider through Hermes's public provider interface.

No MLX packages are injected into Hermes's Python environment. Viewing the library or provider picker does not start a server or change the active provider. The model-provider half requires in-process plugin execution; it cannot return its live SDK client through strict host-process isolation.

Install

Before catalog acceptance, install the repository URL:

hermes plugins install https://github.com/vitaliy-blprnt/hermes-plugin-omlx
hermes plugins enable omlx

Once the catalog submission is merged, hermes plugins install omlx is also available.
Enable its Desktop half under Capabilities → Plugins, then open oMLX in plugin settings.
Both the Python backend and Desktop toggle must be enabled. If an already-running gateway has not loaded the new backend, restart it from an external terminal when ongoing work has finished.

  1. Choose Install.
  2. Download an MLX repository by its exact owner/model ID.
  3. Choose Turn on, then Use beside a downloaded model.

Use runs a short real generation check before saving that profile's default model. It does not rewrite an already-running chat's model. Select omlx and the downloaded model in that chat's usual picker when needed.
After explicit enablement, an inference request can restart the owned runtime. Turn off disables that behavior. Registration, credential resolution, catalog listing and client construction never start it.

Downloads and models

Completion counts the whole snapshot, including cached files. Already downloaded is the cached baseline; received and network rate are separate transfer counters. Buffered file reconstruction is not treated as network speed, and no reconstruction-based ETA is shown.

Cancel reaps the worker before reporting cancellation and keeps hidden partial files for retry. Cancellation is no longer accepted once atomic snapshot publication begins. There is no pause button or automatic cache purge.

Deletion refuses traversal, symlinks, incomplete/unmanaged snapshots, the active default model, and deletion while the owned server is running. No download, activation or uninstall operation deletes a model automatically.

Runtime, models, logs and resumable cache live under the active profile's durable plugin-data/omlx/ directory. Plugin updates do not replace these files. The plugin does not adopt, move or delete models/runtime installations from the earlier core implementation or a separately installed oMLX.

Security and operational disclosure

  • Network: setup fetches the pinned DMG from the official jundot/omlx GitHub release, verifies its committed SHA-256, and downloads public model files from Hugging Face/CDN endpoints. Normal inference is loopback-only. No telemetry is added.
  • Native commands: installation invokes macOS hdiutil (read-only mount/detach) and ditto. Runtime/download operations invoke the private bundle's omlx-cli and omlx-cluster-python; no shell command interpolation is used. Their HOME is private to the plugin so upstream helper scripts are not written into the operator's home.
  • Background processes: the server is a long-running native process; install/download/activation are bounded background jobs. Model downloads can consume substantial bandwidth/disk; loaded models consume unified memory.
  • Ownership: the server binds only 127.0.0.1 at a deterministic per-profile port. The plugin checks process birth identity, executable and owner identity before lifecycle operations. Port conflicts and unverifiable ownership fail closed; it never adopts or kills an arbitrary listener.
  • Credentials: a local-only API key is generated on first start and stored through Hermes's credential API, read with profile-scoped get_secret. It is not written into plugin data or returned by the REST API. The optional masked local_api_key setting names HERMES_OMLX_LOCAL_API_KEY; changing it requires stopping/restarting the server. No external account key is required.
  • Reads/writes: Hermes's documented APIs read active-profile config, persist an explicitly selected default and store the plugin's local secret. Model/runtime data stays in the plugin data directory; the plugin never reads credential files, browser profiles or another tool's login directly.
  • HTTP: local credentials are attached only after verifying the owned runtime. The client rejects another profile/origin and disables HTTP proxy inheritance and redirects. A startup failure is raised at request time, outside core's provider-constructor fallback path.

Only public Hermes extension surfaces are used: ProviderProfile, register_provider, profile-safe plugin storage/config/model-selection helpers, the authenticated plugin backend mount and @hermes/plugin-sdk. No core monkeypatching, internal renderer imports, self-updating plugin code or committed executables.

Development

From a Hermes development environment with pytest available:

python -m pytest
hermes plugins validate "$PWD" --install-deps
node --check desktop/plugin.js

The test fixtures use isolated homes and disposable models/processes, never an operator's library. Native-runtime and rendered-UI checks are separate from mocked tests; see verification notes for the published version's actual evidence and remaining limits.

Provenance

The runtime/download behavior is derived from the unmerged managed-oMLX work in NousResearch/hermes-agent PR #147052, closed in favor of this standalone plugin. The original copyright notice is retained in LICENSE. oMLX itself is developed by its upstream authors and fetched only during explicit setup; it is not bundled here.

README from the reviewed commit eb748ca; it updates when the author re-pins.