VIVA LLM Skill

This skill enables Claude Code to configure and use the VIVA LLM plugin — a multi-provider chat and voice interface embedded in Obsidian.

Overview

VIVA LLM turns Obsidian into a multi-provider AI workstation. It supports text chat, realtime voice calls (OpenAI Realtime, Gemini Live), an integrated terminal, custom assistants, vault-aware tool calling, MCP server integration, screen capture, and agent mode with recursive tool execution.

Supported Providers

Provider	Default Model	Endpoint
OpenAI	`gpt-5-nano`	`https://api.openai.com/v1`
Gemini	`gemini-3.1-flash-live-preview`	`https://generativelanguage.googleapis.com/v1beta/openai`
Anthropic	`claude-3-7-sonnet-20250219`	`https://api.anthropic.com/v1`
xAI	`grok-4-fast-non-reasoning`	`https://api.x.ai/v1`
Mistral AI	`mixtral-8x7b`	`https://api.mistral.ai/v1`
DeepSeek	`deepseek-llm`	`https://api.deepseek.com/v1`
Ollama	`qwen3:0.6b`	`http://localhost:11434/v1`
Cohere	`command`	`https://api.cohere.ai/v1`

Commands

Command ID	Name	Description
`choose-model-and-provider`	Choose model and provider	Switch active LLM provider and model
`complete-chat-response`	Complete chat response	Force completion of a streaming response
`open-voice-call`	Open voice call panel	Open the realtime voice interface
`open-integrated-terminal`	Open integrated terminal	Launch the built-in terminal pane
`speak-chat`	Speak chat	Read the latest response aloud via TTS
`edit-selection`	Edit selection	Apply LLM to the selected text
`send-active-note-context-to-terminal`	Send active note context to terminal	Pipe current note into the terminal session
`generate-title`	Generate title	Auto-generate a chat title
`analyze-conversation`	Analyze conversation	Run conversation analysis
`reverse-roles`	Reverse roles	Swap user/assistant roles
`create-terminal-graph-session`	Create terminal graph session	Open a graph-linked terminal

Key Settings

Setting	Type	Default	Description
`endpoint`	number	`1`	Active provider index
`defaultmaxTokens`	number	`4096`	Max tokens per response
`agentMode`	boolean	`true`	Enable recursive tool-calling agent
`maxRecursionDepth`	number	`20`	Agent recursion limit
`useYAMLFrontMatter`	boolean	`true`	Read/write YAML frontmatter
`skillsFolder`	string	`_skills`	Vault folder for custom skills
`enableMcpTools`	boolean	`true`	Enable MCP tool servers
`enableScreenCapture`	boolean	`true`	Allow screen capture for vision
`screenCaptureFps`	number	`0.5`	Screen capture frame rate
`forceToolCalling`	boolean	`true`	Force tool use in responses

Voice / Realtime Settings

Setting	Type	Default	Description
`realtimeModel`	string	`gpt-realtime-1.5`	OpenAI realtime model
`geminiLiveModel`	string	`gemini-3.1-flash-live-preview`	Gemini Live model
`realtimeVoice`	string	`cedar`	Voice for realtime sessions
`voiceIdleTimeout`	number	`10`	Seconds before voice auto-stops
`voiceBackendContextEnabled`	boolean	`true`	Send vault context during voice
`voiceBackendContextInterval`	number	`20`	Seconds between context pushes

MCP and Orchestrator Settings

Setting	Type	Default	Description
`mcpAppHostEnabled`	boolean	`true`	MCP app-host mode
`mcpDynamicToolsEnabled`	boolean	`true`	Dynamic MCP tool discovery
`mcpDynamicToolLimit`	number	`128`	Max dynamic tools
`realtimeOrchestratorEnabled`	boolean	`true`	Voice orchestrator
`realtimeOrchestratorEndpoint`	string	`http://localhost:8317/v1`	Orchestrator endpoint

Assistants

VIVA LLM supports custom assistants — named personas with their own system prompt, model, and tool/skill configuration:

{
  "id": "custom-id",
  "name": "My Assistant",
  "description": "A custom assistant for X",
  "systemPrompt": "You are...",
  "modelId": "gpt-5-nano",
  "enableTools": true,
  "enabledToolNames": [],
  "enabledSkillIds": [],
  "skillPreferences": {},
  "createdAt": 1700000000000,
  "updatedAt": 1700000000000
}

The default assistant is "default" and can be switched via currentAssistantId.

Skills System

Skills live in the vault folder specified by skillsFolder (default _skills). Each skill is a markdown file the LLM loads as context. Skills can be enabled/disabled per assistant.

Templates

Two template categories are available:

Selection templates — applied to highlighted text (e.g., "Add emojis", "Auto complete 5 suggestions")
Chat templates — full conversation starters

Templates can be configured via selectionTemplates, chatTemplates, and their command-palette variants CMDselectionTemplates, CMDchatTemplates.

Message Format

Chat messages use a role-based format controlled by messageRoleFormatter:

# role: user
What is the mitral valve?

# role: assistant
The mitral valve is...

Tool System

VIVA LLM exposes vault-aware tools to the LLM, grouped by classification. Tools can be individually enabled/disabled via enabledToolClassifications, disabledToolNames, and toolAutoExecution. Agent mode (agentMode: true) allows the LLM to chain tool calls up to maxRecursionDepth.

Excalidraw Integration

Built-in Excalidraw server management:

Setting	Default	Description
`excalidrawManagedStartupEnabled`	`true`	Auto-start Excalidraw servers
`excalidrawAutoRestart`	`true`	Restart on crash
`excalidrawMaxRestarts`	`3`	Max restart attempts

References

Plugin manifest: viva-llm v2.0.2
Caliber source: /Users/mikhail/Obsidian/vivax/.obsidian/plugins/viva-llm/.caliber/summary.json

viva-llm