plugins.md

Meet2Notes Plugin API v1

Meet2Notes plugins can extend the post-recording pipeline, register selectable transcription, diarization, summary, and embedding providers, contribute models to an existing provider, and provide a RAG vector store. The saved recording and canonical transcript remain owned by Meet2Notes; filters work with temporary, serializable artifacts used by downstream stages.

This page is the user and API reference. Package authors should also read Plugin and provider development; the repository's documentation index explains the role of every document.

Community catalog and installation

The public community plugin catalog is a small, maintainer-controlled list of independently developed plugins. Plugin source code and releases stay in the author's repository. An entry means only that the maintainers chose to make the project discoverable; it is not a security audit, warranty, or endorsement.

Meet2Notes does not currently download or install this catalog. Users must inspect the linked repository and install a chosen package explicitly into the private Meet2Notes .venv, using the exact command documented by that plugin. A typical PyPI installation on Windows is:

.\.venv\Scripts\python.exe -m pip install package-name

For a Git repository it may be:

.\.venv\Scripts\python.exe -m pip install git+https://github.com/author/repository.git

After installation, open Settings -> Plugins, select Rescan installed plugins, review the manifest and permissions, and enable the plugin. Rescan only discovers packages already installed in the environment; it does not contact the catalog, PyPI, or GitHub. To remove a plugin, disable it, uninstall the package with the same Python environment, and rescan.

Plugin authors request a listing through the repository's Community plugin listing issue template. Maintainers may accept, decline, update, or remove an entry. The author remains responsible for releases, support, security fixes, and compatibility declarations.

Package discovery

Plugins are regular Python packages installed into the private Meet2Notes .venv. Declare an entry point in the plugin's pyproject.toml:

[project.entry-points."meet2notes.plugins"]
example = "meet2notes_example.plugin:create_plugin"

Discovery does not enable a third-party plugin automatically. There is no watched plugins/ directory: discovery uses installed Python package metadata from the meet2notes.plugins entry-point group.

For local development on Windows:

.\.venv\Scripts\python.exe -m pip install -e .\examples\plugins\analysis_glossary

Minimal filter

from local_meeting_ai.plugins import (
    HookContext,
    MeetingDocument,
    PluginManifest,
    PluginRegistrar,
)


class ExamplePlugin:
    manifest = PluginManifest(
        id="example.analysis-filter",
        name="Example analysis filter",
        version="1.0.0",
        author="Example contributor",
        description="Changes a derived AI input without editing the transcript.",
        permissions=("read_transcript", "write_derived_artifact"),
    )

    def register(self, registrar: PluginRegistrar) -> None:
        registrar.add_filter("analysis.before", self.transform, priority=50)

    def transform(
        self,
        document: MeetingDocument,
        context: HookContext,
    ) -> MeetingDocument:
        result = document.model_copy(deep=True)
        # Transform result.segments here. Never access SQLite directly.
        result.metadata["processed_by"] = self.manifest.id
        return result


def create_plugin() -> ExamplePlugin:
    return ExamplePlugin()

Callbacks may be synchronous or asynchronous. A filter should return the same artifact type it receives. An action returns None.

Stable hooks

Hook	Kind	Payload
final_transcription.completed	Action	Final job metadata
final_transcription.failed	Action	Final job metadata
final_transcription.cancelled	Action	Final job metadata
diarization.completed	Action	Diarization job metadata
diarization.failed	Action	Diarization job metadata
diarization.cancelled	Action	Diarization job metadata
analysis.before	Filter	MeetingDocument
analysis.after	Filter	AnalysisArtifact
analysis.completed	Action	Analysis job metadata
analysis.failed	Action	Analysis job metadata
analysis.cancelled	Action	Analysis job metadata
pipeline.finished	Action	Pipeline result metadata
rag.vector_store.catalog	Filter	VectorStoreCatalog
rag.vector_store.operation	Filter	VectorStoreOperation

analysis.before is the correct location for translation, redaction, terminology correction, or enrichment before the LLM. Set analysis_language when producing a translated document. The original text in SQLite is unaffected.

rag.vector_store.catalog may append descriptors for destinations such as Chroma or a remote vector database; it must preserve the built-in sqlite entry. For a selected plugin destination, rag.vector_store.operation receives one of rows_for_transcription, replace_transcription, candidates, counts, or clear. A plugin handles only its own store_id and returns a copied operation with handled=True and a serializable result. Search candidates must include the embedding plus the meeting, transcript, chunk, timestamps, text, title, and date needed by the core ranker and provenance UI.

Ordering and failures

Lower priorities run first. Equal priorities are ordered by plugin ID for reproducibility:

registrar.add_filter(
    "analysis.before",
    callback,
    priority=40,
    timeout_seconds=30,
    failure_policy="continue",  # or "fail"
)

Every call records plugin ID/version, hook, job and pipeline identifiers, duration, status, error, and SHA-256 input/output digests. Transcript or summary content is never stored in the plugin execution ledger.

Security rules

• A Python plugin is executable code. Install and enable only packages you trust.
• Do not read Meet2Notes' SQLite database or model directories directly.
• Declare permissions accurately. Network access must never be hidden.
• Do not log transcripts, API keys, tokens, or recording paths.
• Use declarative provider settings for non-secret configuration and the operating-system keyring for secrets.
• Pin dependencies and test against the declared Plugin API version.

Third-party process isolation, signed packages, in-app installation, staged updates, and rollback are tracked in the roadmap. The current JSON catalog remains informational until those protections exist.