hermes-openwhispr

A small, local-first plugin between Hermes and OpenWhispr. Offered as a standalone plugin — not a claim on Hermes core, just an option for people who want it.

Two doors, pick yours.

You dictate and want your words found — try this (8 lines):

brew install --cask openwhispr     # or openwhispr.com/download
# open the app, press Globe (macOS), Ctrl+Win (Windows), or Ctrl+Super / F8 (Linux), say something
npm install -g @openwhispr/cli
openwhispr doctor                  # local bridge reachable
openwhispr --local notes search "what you just said" --limit 5

New to OpenWhispr? Start at their quickstart (docs.openwhispr.com/quickstart), then come back. You maintain Hermes and you're reviewing a proposal:

hermes plugins install owner/repo  # or open hermes://plugin/install?repo=owner/repo&enable=1 on desktop
hermes plugins enable hermes-openwhispr
# then in a session: skill_view("hermes-openwhispr:openwhispr")
# plugin skills load by qualified name only — they are not auto-indexed

One worked session, the whole product:

You: "find what I just dictated about the dentist" Agent: openwhispr --local notes search "dentist" --limit 20 → note id 7, created_at this morning → notes get 7 --transcript → your words, quoted back with id + timestamp.

The pictures below tell the same story visually; the prose after them is the maintainer's case.

Words used once, for everyone: STT means speech-to-text (voice into words). MCP means Model Context Protocol (how assistants read app data). A gateway is any Hermes front door — Discord, Telegram, the terminal, and friends. CLI, TUI, dashboard, and desktop are the four local faces: command line, terminal UI, web panel, native app.

What it does

Reads your OpenWhispr dictation history (transcriptions) so an agent can find what you said, even when you can't remember it.
Reads, searches, and (with your say-so) writes OpenWhispr notes as a handoff surface between agents.
Transcribes local audio with your desktop model, optionally filing it straight to a note.
Manages dictionary words and snippets from the terminal.

What it doesn't do: replace Hermes STT, include assistant replies, or touch the cloud unless you explicitly opt in.

Why this sells itself (not our product, just the facts)

Per-feature model choice across tiers — Dictation Cleanup, Voice Assistant, Translation, Note Formatting, Chat each pick independently from: OpenWhispr Cloud (no setup), Cloud Providers (bring your own key), Local (on-device, fully private), Self-Hosted (your server on your network), Enterprise (your org's cloud account — AWS and Azure real, GCP marked planned in their docs). A live transcription preview can show in a floating window before cleanup (see the dictation shot below).

Auto-learn from corrections is off by default; once you enable it under Settings → Preferences, the words you fix in the target app land in your dictionary by themselves. "Automatic pasting" drops text into the active app, "Keep transcription in clipboard" holds it for manual pasting. Optional file export saves notes and transcripts to disk organized by folder, with "Rebuild all files" — a second on-disk breadcrumb trail agents can read directly. Calendar integrations are all optional; only Google's row promises auto-filled titles and attendees, Microsoft and Apple say "read," and Apple reads Calendar.app. CLI and MCP split cleanly: local CLI free with no key, cloud CLI and assistant connectors (a paid tier, buttons as labeled in the app's Integrations screen).

Under the hood it's bundled C++ runtimes, not a Python stack: whisper.cpp for the Whisper family, sherpa-onnx for Orukeet, Parakeet, Nemotron, and Cohere — no interpreter or runtime to install. Metal acceleration is built in on Apple Silicon with automatic CPU fallback, CUDA/Vulkan are one-click on other platforms, and the default local model is ~672MB with the most accurate offline option at ~2.7GB. Small binaries, honest sizes, GPU when it's there, CPU when it isn't.

One trail across every surface. Discord, every gateway, every interface — CLI, TUI, dashboard, desktop. Because OpenWhispr sits at voice input rather than inside any one app, it's a rolling record of everything you say across all of those trajectories, not just one chat log. Wherever the words came out, the breadcrumb is there.
Scale it to your machine and your appetite, not five fixed sizes. Small computer with thin resources, or light needs? Go small. Want a secretary in the chat interface minting notes alongside Hermes into a shareable location? Go big. Sharing anything is opt-in; staying private is a switch — turn everything off and it all still works on-device. All up to you.

Screenshots

Read top to bottom — each picture follows the claim it proves, with its context right beside it so nothing needs backtracing. Personal details redacted with blur before publishing.

Integrations screen: assistant upsell on top, CLI access below

The eye hits the paid connector card first; the CLI card underneath is the free half: npm install -g @openwhispr/cli, then exactly what the card shows — openwhispr --local notes list with no login. Dictations land as notes automatically, so this is the front door, not an empty room: for anything older than today, openwhispr --local notes search "<phrase>" --limit 20, then notes get <id> --transcript.

Per-feature model picker: five tiers, your call each time

Dictation Cleanup, Voice Assistant, Translation, Note Formatting, Chat — each picks independently from OpenWhispr Cloud, Cloud Providers (own key), Local on-device private, Self-Hosted, or Enterprise.

Auto-learn from corrections (shown after enabling it)

Fix a word in the target app and it joins your dictionary — but only after you flip this switch on. Nothing learns silently.

Clipboard flow plus save-notes-as-files (save path redacted)

"Automatic pasting," "Keep transcription in clipboard," and on-disk Markdown organized by folder with "Rebuild all files" — a second breadcrumb trail agents can read directly.

Optional calendars and paywalled API access (account email redacted)

Only Google is connected here, which is the row that auto-fills meeting titles and attendees. API access itself sits behind the paid plan.

A different picker: dictation engines, four tiers, four local vendors

Not the same dialog as above: Dictation, Note Recording, Audio Upload across Cloud, own-key, Local, Self-Hosted — no Enterprise tier here. Local vendors Oruk, OpenAI, NVIDIA, Cohere, with Cohere Transcribe 2B active on-device — plus the live-transcription floating preview.

Dictation Cleanup download list: six vendors, honest sizes

Qwen, Mistral, Meta Llama, OpenAI, Gemma, Liquid AI. These are downloads, not installs — Gemma 4 31B shows a Download button at 19.6GB, and the list continues below the crop. Pin a different model per task or run everything on one — local, cloud, self-hosted, or enterprise, decided per feature, not per app.

The point in one paragraph

No Python dependencies and no changes to Hermes itself — though be honest about the price of admission: Node 20, one global CLI package, and the desktop app running (plus ffmpeg on PATH for large cloud uploads). What you get is your dictations — not assistant replies — showing up in every space you interact in, with every agent you talk to.

A win for both sides — and everyone in between

OpenWhispr meets a crowd of local-first users who would never otherwise try it. Free local gets them in; the paid cloud, MCP connectors, and API are one step away when they want more — plus an agent bootstrap flow that mints scoped keys with a single email code.
Hermes and Nous Research get the option of a local default that remembers: every user carrying a rolling record of what they said, reachable from any surface, on top of transcription they already trust.
Users stop choosing between a bill and a memory. Start free and private; pay only if the cloud earns it; never lose what you said. Tie in email, keep notes, transcribe audio — small pieces, and together nearly a powerhouse.

Privacy is yours to tune

Granular, both directions, your call throughout:

Local: free, no login, audio never leaves your machine. Requires the OpenWhispr desktop app running (desktop bridge on loopback; the live port is recorded in the CLI's bridge file). This is the default and works without any key.
Cloud: opt-in only. Needs a Pro or Business plan plus openwhispr auth login (the key lives in the CLI's own config, not in Hermes .env). Cloud transcription is beta with no SLA; files over 4 MB are chunked client-side and need ffmpeg on PATH.
Reads vs writes: reads and searches are always safe. Creates, updates, and deletes (notes, transcriptions, dictionary, audio) only happen with your explicit consent, every time.
Dictionary: a hint to the model, not an override. Auto-learn from corrections (Settings → Preferences) is off unless you turn it on. Export it before moving machines.

Disclosures (catalog review)

Shell-outs to openwhispr * through the terminal tool only.
Local network: loopback desktop bridge when using --local. The app writes a one-time bearer token to its own bridge file (mode 0600) at startup and the CLI reads it automatically — loopback is authenticated, not open.
Remote network: api.openwhispr.com only when you opt into --remote.
The agent never opens credential files itself: the vendor CLI reads its own bridge token (~/.openwhispr/cli-bridge.json) and ~/.openwhispr/cli-config.json (both 0600). Key setup includes the agent bootstrap flow (/auth/email-code, a 6-digit code the user pastes, POST /keys/create for a scoped key) — always with consent, key stays in the CLI config, never in Hermes .env. Note transcriptions:delete has no desktop-app checkbox, so a delete-capable key can only be born through the API.
No daemons of ours, no self-updater, no telemetry. (The desktop app itself must be running for local mode — that's theirs, not ours.)
No core overrides: no patching of Hermes functions, modules, or stores. If a needed hook doesn't exist, that's an upstream feature request.

Attribution

OpenWhispr makes the dictation app, the CLI (@openwhispr/cli), the cloud API, and the docs at docs.openwhispr.com this plugin leans on. All screenshots are of their application; all product names and marks are theirs. This plugin is an independent community pitch, and OpenWhispr should know it's being recommended to Nous Research as an option — with appreciation, and without claiming their endorsement.
Hermes Agent by Nous Research is the platform this plugin extends, through public plugin surfaces only. This is a proposal for their consideration, not a directive — adoption as a first-class option is theirs to decide.
This plugin itself is community work, published under MIT, with no affiliation to either party.

Layout

hermes-openwhispr/
├── plugin.yaml
├── __init__.py                      # register_skill(), skill-carrier
├── README.md
├── LICENSE
├── skills/openwhispr/SKILL.md
├── docs/screenshots/01-07*.png
└── tests/test_plugin.py

Install

Command line all the way down — no App Store involved:

brew install --cask openwhispr     # desktop app (verified in the brew catalog)
npm install -g @openwhispr/cli     # agent CLI (pnpm or bun work too)

Alternatives straight from the source: openwhispr.com/download for the signed builds (macOS, Windows, Linux, iOS), GitHub releases for the .dmg files (Apple Silicon and Intel), or git clone https://github.com/OpenWhispr/openwhispr.git plus npm install / npm run dev if you build from source. OpenWhispr itself is MIT-licensed open source. Leave the app running — the local bridge is what the agent talks to.

Smoke test

npm install -g @openwhispr/cli
openwhispr doctor                      # exit 0, local bridge reachable
openwhispr --local notes search "hello" --limit 5
hermes plugins install owner/repo && hermes plugins enable hermes-openwhispr
# then in a session: skill_view("hermes-openwhispr:openwhispr")
# (plugin skills load by qualified name; they are not auto-indexed)
python -m pytest tests/ -q             # here

The Hermes-side skill test (tests/skills/test_openwhispr_skill.py) belongs to a hermes-agent PR, once proposed — not to this repo.

Proposal path

Pitched as a standalone plugin. Hermes's own guide keeps third-party product integrations in standalone repos rather than the core tree, so any ship-bundle conversation would be a separate, optional ask — no presumption either way. This stands alone regardless.