OCP Pipeline

September 2, 2026 · View on GitHub

The OCP (Open Common Play) pipeline plugin turns media utterances into playback actions in OVOS. Utterances such as "play metallica", "put on some movie", "pause", or "next" pass through this plugin as an intent pipeline stage. On a match, the plugin searches registered media skills over the message bus and drives playback through the OCP API.

The plugin exposes two entry points under the opm.pipeline group:

  • ovos-ocp-pipeline-plugin (OCPPipelineMatcher), the main pipeline. It matches play, open, media_stop, next, prev, pause, resume, save_game, and load_game intents at high, medium, or low confidence.
  • ovos-ocp-pipeline-plugin-legacy (MycroftCPSLegacyPipeline), a bridge for older Mycroft CommonPlay skills, using the play:query / play:query.response / play:start handshake.

OCPPipelineMatcher also classifies media type, such as music, movie, or podcast. It extracts named entities with ahocorasick-ner, using per-language vocabulary files and optional user-supplied entity CSVs (the entity_csvs config option). It tracks player state per session through an OCPPlayerProxy, kept in sync with ovos.common_play.status and track.state bus events.

Architecture

image

image

Media classification

Media-type classification, the rich provider-ready descriptive Signals, and the content filter are all delegated to the standalone ovos-media-classifier. The pipeline builds the classifier's two context inputs from its own state — the per-session now-playing PlayerStatus and the skill-registered entities (ner_list) — so relative control follow-ups ("next", "pause", "something else") and entity routing work, and it forwards the classifier's lossless mediavocab.Signals (medium / playback_type / content_genres / content_form / programme_format / variant_kind / accessibility / picture_format) to the MediaProviders alongside the legacy media_type.

With no extra configuration the lean keyword (.voc) backend is used — the zero-ML-dependency floor (all heavy backends OFF, online OFF, adult content blocked). The classifier is configured under the pipeline config (or a nested media_classifier block); every key is optional.

Backend selection

KeyDefaultDescription
media_classifier_pluginunsetname of an external opm.media.classifier entry-point plugin to load
media_classifier_onnx_modelunsetpath to an opt-in ONNX trained bundle (requires the [onnx] extra)
media_classifier_embedding_routerunsetpath to a learned embedding-router bundle (requires the [onnx] extra)
media_classifier_embedding_router_hybridtruerun the router as a keyword+router hybrid (false = router standalone)

On any failure (missing extra, bad bundle, unknown plugin) the classifier falls back to the keyword backend, so the zero-ML default is always preserved.

Gazetteer / entity library (embedding-router backend only)

KeyDefaultDescription
media_classifier_gazetteertrueinject the bundled offline gazetteer of common real titles so bare titles route without a network call
media_classifier_gazetteer_sizeclassifier defaultcap on titles per media type
media_classifier_entity_libraryunset{label: [titles]} of the user's own media library, injected at runtime (no retraining)

Online metadata layer (embedding-router hybrid only)

KeyDefaultDescription
media_classifier_online_metadatarrfalseopt into the online metadatarr last-resort layer (adds latency)
media_classifier_online_timeout4.0per-request timeout, seconds
media_classifier_online_min_confidence0.5minimum confidence to accept an online answer

Content filter (applied at routing)

Blocked content (adult by default) is never routed to providers.

KeyDefaultDescription
allow_adult_contentfalsetop-level convenience flag; true lifts the default adult block
media_content_filter.enabledtruemaster switch for the filter
media_content_filter.blocked_genres["adult"]genres to block
media_content_filter.blocked_media_types[]media types to block

Install

pip install ovos-ocp-pipeline-plugin

Usage

OVOS core loads this plugin automatically once installed, through the opm.pipeline entry-point group. Enable it in mycroft.conf under the intents section:

{
  "intents": {
    "ovos-ocp-pipeline-plugin": {
      "entity_csvs": []
    }
  }
}

The plugin reads the pipeline's intents config block, then falls back to a legacy OCP config block for backward compatibility.

Media providers

Besides broadcasting a search over the message bus to OCP skills, the pipeline can query MediaProvider plugins in the same process. A provider is a plain catalog: it takes a parsed request and returns candidate releases, with no bus round trip and no skill to run.

Install the whole published set with the providers extra, or pick the ones you want one at a time:

pip install ovos-ocp-pipeline-plugin[providers]
pip install ovos-media-provider-somafm ovos-media-provider-local

Nothing else is needed to register them. Each plugin declares an opm.media.provider entry point, and the pipeline instantiates every installed provider at startup. A provider that needs credentials or a reachable server, such as Spotify or Music Assistant, loads anyway and answers nothing until you configure it.

Per-provider settings live in the OCP pipeline's own config block — mycroft.confintentsovos-ocp-pipeline-pluginmedia_providers — keyed by the provider name as it appears in its entry point. A top-level media_providers block is honoured only when the pipeline config carries no media_providers key at all, so keep the settings in the pipeline block. Set enabled to false on a provider to skip it entirely:

{
  "media_providers": {
    "local": {"paths": ["/home/ovos/Music"]},
    "spotify": {"enabled": false}
  }
}

To turn the in-process search off altogether and go back to bus skills alone, set enabled to false on a media_providers block inside the pipeline's own config. Be aware that a media_providers block in the pipeline config replaces the top-level one as the source of per-provider settings, so keep provider settings in one place or the other:

{
  "intents": {
    "ovos-ocp-pipeline-plugin": {
      "media_providers": {"enabled": false}
    }
  }
}

Both searches run for every query that does not name a specific skill, and they run at the same time, so a search costs the slower of the two rather than both added together. Their results are pooled and ranked together. A provider answer never displaces a skill answer: the only thing dropped is a provider entry pointing at a URI a skill already offered.

A provider may declare the media types it serves. When a request names a type, only the providers that claim that type are asked, and a provider that claims nothing is asked but ranked last: its answer plays when nothing else answered, and never wins against a provider or skill that claims the type. A request that names no type asks everything and ranks everything on merit.

License

Apache-2.0