Voice Mode

Purpose

Voice-driven invocation of Maxim's full capability surface. Wraps mbailey/voicemode for STT (Whisper) + TTS (Kokoro) and adds office-aware intent routing, hotword config, and decision capture.

Maxim is the only multi-agent framework with native voice-driven executive office routing.

When To Invoke

• User runs /mxm-voice
• User uploads audio that needs transcription + dispatch
• Voice utterance contains "Maxim, ..." or "Hey Maxim, ..." trigger phrase
• Hands-free workflows (commute, accessibility, eyes-free debugging)

Architecture

This skill DOES NOT do STT or TTS itself. It depends on two layers:

1. mbailey/voicemode MCP — handles Whisper transcription + Kokoro speech synthesis
2. mxm-voice MCP — Maxim's hotword config + intent routing + decision capture

The flow:

[User speaks] → [voicemode Whisper STT]
                      ↓
              [transcribed text]
                      ↓
       [mxm-voice.voice_invoke_command]
                      ↓
       Hotword match? ─── YES ──→ [dispatch slash command]
              ↓
              NO
              ↓
       [/mxm-route <utterance>]  → [executive-router classifies intent]
                      ↓
              [skill execution]
                      ↓
              [text response]
                      ↓
       [mxm-voice.voice_brief_office or formatted response]
                      ↓
              [voicemode Kokoro TTS]
                      ↓
       [User hears response]

When VoiceMode Is Not Installed

/mxm-voice prints install instructions and exits gracefully:

🗣️ VoiceMode not detected. Install with:
   claude plugin marketplace add mbailey/voicemode
   claude plugin install voicemode@voicemode
   /voicemode:install

The mxm-voice MCP server still runs (provides voice_status, hotword config validation, decision capture from typed transcripts) — only the actual STT/TTS requires voicemode.

Hotword Configuration

User-editable file: config/voice-phrases.yml

phrases:
  - pattern: "maxim run my project health check"
    dispatch: "/mxm-health"
  - pattern: "maxim draft a landing page *"
    dispatch: "/mxm-design landing page $1"
  - pattern: "maxim research * tell me how it improves *"
    dispatch: "/mxm-cino research $1 for $2"

Pattern types: plain string (substring match), wildcards (* captured as $1, $2), regex (regex:<pattern>).

User adds/removes/renames phrases freely. Config reloads on every voice invocation — no restart.

Default Hotwords (ship with Maxim)

Spoken	Dispatched
"Maxim, morning CEO cycle"	/mxm-ceo-morning
"Maxim, overnight CEO cycle"	/mxm-ceo-overnight
"Maxim, portfolio status"	/mxm-portfolio
"Maxim, project status"	/mxm-status
"Maxim, run project health check"	/mxm-health
"Maxim, run wiki ingest"	/mxm-wiki ingest
"Maxim, query wiki *"	/mxm-wiki query $1
"Maxim, audit the wiki"	/mxm-wiki lint
"Maxim, review unverified pages"	/mxm-wiki explore
"Maxim, recall *"	/mxm-recall $1
"Maxim, remember *"	/mxm-remember $1
"Maxim, brief me on the *"	voice_brief_office($1)
"Maxim, draft a landing page *"	/mxm-design landing page $1
"Maxim, research * tell me how it improves *"	/mxm-cino research $1 for $2

Behavioral Notes

• Voice utterances bypass the typed CLAUDE.md preamble — keep responses concise (TTS-friendly)
• Sensitive operations (security, compliance, payments, regulated projects) STILL fire CSO auto-loop — voice is not a bypass
• Voice-captured decisions go to .claude-sessions-memory/decision-log.md with timestamp + transcript
• Operator profile picks up speech patterns over time (preferred command verbs, common requests)

Output Format

Voice Invocation:
Heard: "<transcribed utterance>"
Match: HOTWORD | EXECUTIVE_ROUTER | UNCLASSIFIED
Dispatched: <slash command or fallback>
TTS Response: <spoken text — concise>
Captured to memory: YES (decision tag) | NO
Confidence: 🟢 HIGH (clean transcription, hotword match) | 🟡 MEDIUM (router classified) | 🔴 LOW (unclassified, asked for clarification)

Handoff

• Hotword matched → dispatch slash command directly
• No match → /mxm-route <utterance> for intent classification
• Compliance-sensitive utterance → CSO auto-loop fires before any action
• Decision dictation → voice_capture_decision writes to session memory

Maxim Behavioral Layer

• Confidence tagging: 🟢 HIGH on hotword match, 🟡 MEDIUM on router classification, 🔴 LOW on unclassified
• Brevity: voice responses are SHORTER than typed responses (TTS attention budget)
• Anti-misfire: refuses to act on partial/garbled transcriptions (asks for re-utterance)

Source References

• mcp/mxm-voice/ — MCP server implementation
• config/voice-phrases.yml — user-editable hotword config
• .claude/commands/mxm-voice.md — slash command entry point
• mbailey/voicemode — upstream STT/TTS dependency

---

Copyright (c) 2026 iSystematic Inc. Maxim product. BSL 1.1.