Hermes Voice Assistant

Hermes is a Real-Time, interactive voice interface for your Obsidian vault.

Powered by Google's Gemini X Native Audio API, it allows you to talk directly to your notes, perform file operations, and search the web—all via a low-latency voice channel.

What?

It's like talking to OpenAI's ChatGPT voice interface, but with direct access to your vault, so you can structure, organize your thoughts and notes, without touching a key. When next time you sit down to it, you'll have all the output to go on, in a structured way.

To this day, I have not found a tool that could do that, so I sat down vibe coding. 24 hours later I am here, having a talking voice agent at my disposal, that uses my google cloud free tier.

"I just want to talk to my vault, while I walk, like though a vaulty-talkie?"

Demo

https://www.youtube.com/watch?v=FcX2EzMf8GY

Why?

What the problem for me with OpenAI's been, was that I never find stuff I talked about. The goal is not to produce/distill something, rather to just a stream of data.

AI is good at generating text. Well, soulless, slop, sometimes. Where is you in there? You are the curator. The builder. Who keeps the focus. Who cycles thoughts.

With your human sense of reality can we only flight the slop.

So yeah - those long walks I had talking about stuff in the forest, might as well have productive outputs from now on!

I bet you're here for similar reasons.

"Most of our work happens in the tram"

• Jeszy about our profession

I guess there are other knowledge workers like that out there.

If you think this is useful for you, or this is what you've been waiting for, this is your lucky day too.

Disclaimer

For your safety, it's 98% Vibe coded over one weekend. I just tested it on my live vault. It looks pretty ok, does not accidentally delete all the files...

I certainly DO recommend some safety measures/backups before using it! Something like using a git repository, syncthing file history, or whatever your weapon of choice is.

I most certainly do not take any responsibility for giving accidental commands for deleting ALL the files in your repo!

Privacy

Also, be mindful, you're giving access to your notes to a tech giant.

But as some wise person said, post Orwell: "The price of privacy is the loss of convenience."

So yeah, Google's LLM Will read your notes. That's why you can talk to them.

I looked into self hosted solutions, but our home hardware is just not there yet.

Until then, the world is going by.

✨ Features

• Gemini Live Integration: Natural, human-like voice conversations with real-time audio streaming
• Vault Awareness: Hermes can read your notes, list directories, and understand the context of your current folder
• Smart Tool Calling:
  • read_file / create_file / update_file: Full file lifecycle management
  • edit_file: Targeted line-based modifications
  • search_keyword / search_regexp: Global vault searching
  • internet_search: Real-time web grounding via Google Search
  • generate_image_from_context: AI-powered image generation
  • topic_switch: Automatic conversation archiving
• Automatic Archiving: Using the topic_switch tool, Hermes can summarize segments of your conversation and save them as markdown notes in your vault

🚀 Installation

BRAT (Beta, Recommended)

Currently available in Beta via BRAT:

1. Install BRAT from the Community Plugins store.
2. In BRAT settings, click Add Beta plugin.
3. Enter: https://github.com/symunona/obsidian-hermes
4. Enable Hermes Voice Assistant in your Community Plugins list.

Manual Installation (why would you do that?)

1. Download main.js, manifest.json, and styles.css - under releases
2. Create a folder named hermes-voice-assistant in your vault's .obsidian/plugins/ directory
3. Move the downloaded files into that folder
4. Reload Obsidian and enable Hermes Voice Assistant in the Community Plugins settings

B: you can just check it out from the git repo if you're reading this.

🛠 Developer Reference

Project Structure

• main.ts: The entry point for the Obsidian plugin
• HermesMainViewObsidian.tsx: The Obsidian Workspace Leaf that hosts the React application
• App.tsx: The root React component managing state and the voice session
• services/voiceInterface.ts: Core logic for managing the @google/genai Live session, audio encoding/decoding, and tool execution
• services/commands.ts: Tool registry and execution engine
• tools/: Individual tool definitions (declarations and execution logic)
• components/: UI components (Chat, Tool Results, Settings, Kernel Log)
• persistence/: Settings and data persistence layer
• utils/: Utility functions for audio processing, environment detection, etc.

Started working also on an Obsidian Independent version, like the app that can just run standalone. Maybe the next free weekend I have.

Development Commands

# Install dependencies
pnpm install

# Start the build watcher (for development)
pnpm run dev

# Build for production
pnpm run build

# Build CSS only
pnpm run build-css

# Watch CSS changes
pnpm run watch-css

# Serve standalone version
pnpm run serve

# Development with standalone
pnpm run dev-standalone

# Build standalone version
pnpm run build-standalone

Available Tools

The plugin includes 20+ tools for vault operations:

File Management:

• read_file: Read file contents
• create_file: Create new files
• update_file: Update entire file content
• edit_file: Line-based editing
• delete_file: Delete files
• rename_file: Rename files
• move_file: Move files between directories

Directory Operations:

• list_directory: List directory contents
• list_vault_files: List all vault files with filtering
• get_folder_tree: Get folder structure
• dirlist: Quick directory listing
• create_directory: Create new directories

Search & Replace:

• search_keyword: Search for text patterns
• search_regexp: Regex-based search
• search_replace_file: Search and replace in single file
• search_replace_global: Global search and replace

Advanced Features:

• internet_search: Web search via Google
• generate_image_from_context: AI image generation
• topic_switch: Archive conversation segments
• end_conversation: Graceful session termination

Adding New Tools

To add a new capability to Hermes:

1. Create a new file in tools/[tool_name].ts
2. Define a declaration (OpenAPI schema) and an execute function
3. Add an entry to the TOOLS registry in services/commands.ts
4. Add the tool's usage instructions to utils/defaultPrompt.ts

I was thinking of adding the capacity to pull in ANY MCP, that'd make sense. But that's not the first weekend.

Tech Stack

• AI: @google/genai (Gemini 2.5 Flash Native Audio)
• UI: React 19, Tailwind CSS
• Markdown: marked for md rendering
• Build: esbuild for plugin, Vite for standalone (WIP)
• Audio: Web Audio API for real-time audio processing
• Package Manager: pnpm

⚙️ Commands

Obsidian Commands

Speak naturally to Hermes to:

• "Read my notes" → Lists current directory
• "Create a new note called meeting-notes" → Creates file
• "Search for TODO items" → Searches vault
• "What's the weather today?" → Web search
• "Archive this conversation" → Saves chat history

Fine tune it under settings!

🔧 Settings

• Get your Google API key at: https://aistudio.google.com/api-keys

Behavior Settings

• Voice Persona: Select voice character
• Custom Context: Add persistent context to all sessions
• System Instructions: Override default AI behavior
• Chat History Folder: Configure archive location

It'd be good to have per folder instructions, right? PRs welcome.

🐛 Troubleshooting

Common Issues

1. Connection Issues: Check network connectivity and API quota - be mindful of your quota! Have no clue when you'll start paying. They seem to be generous though. You might have to activate the billing account to work flawlessly for longer sessions.
2. Large Vaults: Use list_vault_files with limits for better performance - I tried working around by doing smart chunking, but it's not a fully solved problem.

Mobile support: I did not test. Will do.

⚖️ License

See LICENSE

Hermes: Bridging the gap between thought and file.