AI Assistant

The AI assistant plugin provides a full conversational AI experience in Mumble. It supports text chat, voice input and output, emotion tracking, karma feedback, and persistent memory — all powered by OpenRouter.

OpenRouter Integration

Feature	Details
Provider	OpenRouter (openrouter.ai/api/v1)
API key	Set via api_key config or api_key_env_var env var.
Chat model	Configurable (default: openai/gpt-4o-mini), switchable at runtime with !chatmodel.
Server-side tools	web_search, datetime_tool, web_fetch_tool — processed by OpenRouter.
Reasoning	Configurable effort (low/medium/high), thinking block stripping.
Context compression	OpenRouter can trim older messages to stay within context.

Client-Side Tools

The bot executes these 15 tools locally when the AI requests them:

Tool	Description
get_server_state	Connected users, channels, bot state.
get_now_playing	Current track info.
get_channel_messages	Recent channel chat.
get_user_notes	Memory notes for a user.
get_conversation_history	Chat history.
save_memory	Save a note.
delete_memory	Delete a note.
get_bot_logs	Read runtime logs.
get_bot_source	Read bot source (if self_inspect=true).
get_sfx_clips	List soundboard clips (if soundboard_tools=true).
play_sfx	Play a soundboard clip.
play_media	Queue YouTube/SoundCloud.
play_radio	Play a radio stream.
control_playback	Stop/skip/pause/resume.
get_karma_feedback	Read karma feedback.

Text-to-Speech (TTS)

• Synthesizer: OpenRouterSynthesizer using OpenRouter’s TTS endpoint.
• Voice model: google/gemini-3.1-flash-tts-preview (default).
• Voices: 30 Gemini voices available (default: Kore).
• Inline SFX: The AI can embed [[sfx: clip_name]] tags in responses. The TTS pipeline pauses, plays the clip, and continues.
• Toggle: !chatvoice enables or disables voice output.

Voice Input (STT)

• Transcriber: OpenRouterTranscriber using openai/whisper-1.
• VoiceListener captures PCM audio from Mumble, detects silence boundaries, and sends segments for transcription.

Command	Description
!speak	Start listening.
!speakcancel	Cancel listening.

Emotion System

The AI tracks emotional state across 10 states:

neutral, happy, excited, amused, touched, annoyed, salty, bored, frustrated, defensive

The AI emits [[mood: state]] and [[user_sentiment: ...]] tags (stripped before display). Mood follows a valence/arousal model with decay toward neutral over time.

Abuse detection: Rapid command spam triggers a penalty that pushes the mood toward negative states.

!feed <item> — The AI judges an item on a -10 to +10 scale and adjusts its mood accordingly.

Karma System

Command	Description
!chatkarma +1/-1 [note]	Rate the last AI response.
!chatkarma 0	Undo a rating.
!chatkarmastats	View karma statistics.

Karma feedback influences the emotion system.

Memory

The AI maintains persistent per-user and global notes (stored as JSON). It can save and delete notes via its tools.

Command	Description
!chatnotes	View your saved notes.
!chatforget	Delete all your notes.

Personas and Voice Profiles

• Personas are system prompt files that shape the AI’s personality. Switch at runtime with !chatpersona.
• Voice profiles are speech prompt files that adjust tone and style. Switch with !chatvoiceprofile.