Emotion Machine Voice Plugin for OpenClaw

This plugin connects OpenClaw to Emotion Machine’s voice workspace API so long‑running tasks can update hot_context and be spoken by the Fast Brain.

What It Does

• Provides an em_voice_task tool for logging task lifecycle events to EM /voice/workspace
• Exposes gateway RPC methods (em-voice.start, em-voice.done, etc.)
• Optionally syncs contacts from USER.md into EM relationships

Install

# From npm
openclaw plugins install @emotion-machine/em-voice

# From local folder (development)
openclaw plugins install ./openclaw-plugin

Restart the OpenClaw gateway after install or config changes.

Publish

cd openclaw-plugin
npm run clean
npm run build
npm publish

Notes:

• Package includes compiled dist/ output via prepare/prepublishOnly.
• Scoped package is configured for public publish (publishConfig.access=public).

Configure

Add to your OpenClaw config:

{
  plugins: {
    entries: {
      "em-voice": {
        enabled: true,
        config: {
          em: {
            apiKey: "em_live_...",
            companionId: "comp_abc123...",
            workspaceApiToken: "voice_ws_...",
            // baseUrl: "https://api.emotionmachine.ai"
          },
          workspace: {
            path: "~/.openclaw/workspace"
          }
        }
      }
    }
  }
}

Notes:

• workspaceApiToken must match VOICE_WORKSPACE_API_TOKEN on the EM server.
• relationshipId is required in production when calling em_voice_task.
• If you want a single global hot_context for non‑phone tasks, set em.defaultRelationshipId and omit relationshipId in those tool calls.

Usage

Tool call from the agent

{
  "action": "done",
  "taskId": "abc123",
  "relationshipId": "rel_...",
  "result": "Email draft ready."
}

Gateway RPC (manual / debugging)

openclaw rpc em-voice.start --params '{"relationshipId":"rel_...","taskId":"abc123","query":"Draft email"}'
openclaw rpc em-voice.done --params '{"relationshipId":"rel_...","taskId":"abc123","result":"Draft ready"}'
openclaw rpc em-voice.status --params '{"relationshipId":"rel_...","taskId":"abc123"}'

One-Command Setup

openclaw em-voice init

This will:

• Open a browser to authorize Emotion Machine (OAuth + PKCE)
• Derive a fast-brain prompt (and pick a voice) from your local SOUL.md / IDENTITY.md
• Create a default global relationship
• Print a single openclaw config set ... command to apply
• Print the Twilio inbound webhook URL for this companion

The init is idempotent: it will reuse an existing companion if found in config or workspace. Use --reinstall to force a new companion. You will be prompted to apply the config and restart the gateway automatically.

The companion/relationship IDs are also stored in: ~/.openclaw/workspace/.em/em-voice.json

If the browser cannot reach your CLI, you can copy the code from the page and paste it back into the terminal.

For local dev:

openclaw em-voice init --base-url http://localhost:8100 --auth-url http://localhost:3000

Flags:

• --auth-url to override the dashboard origin.
• --manual to skip the localhost callback and paste the code manually.
• --base-url is only needed for dev/staging; default is production (https://api.emotionmachine.ai).
• --companion-id to reuse an existing companion.
• --reinstall to force re-create the companion.
• --voice-name, --preset, --stt-provider, --tts-provider to control the voice pipeline.
• --caller-phone to create phone:<digits> relationship for your own phone.
• --twilio-account-sid, --twilio-auth-token, --twilio-phone-number to enable BYO Twilio.

What Gets Synced

| OpenClaw File | → | EM Entity | |---------------|---|-----------| | SOUL.md / IDENTITY.md | → | companion.config.system_prompt | | USER.md (contacts) | → | relationship.profile | | memory/contacts.md | → | relationship.profile |

Profile Structure in EM

{
  "user": {
    "name": "Sample User",
    "email": "[email protected]",
    "phone": "+1 555 010 0000",
    "timezone": "America/Los_Angeles"
  },
  "relationship": {
    "role": "employee_1",
    "trust_level": "full",
    "source": "USER.md",
    "synced_at": "2026-02-06T10:00:00Z"
  }
}

How It Works

1. Webhook received → Task logged to hot_context.md
2. Fast brain → Reads context, generates immediate acknowledgment (~200ms)
3. Response sent → User hears "Got it, checking..." via TTS
4. Slow brain → OpenClaw agent processes the full request (async)
5. Callback → Full result POSTed back to your voice app
6. TTS → Your app speaks the complete answer

Race Condition Safety

The hot_context.md file uses an append-only event log pattern:

• No locking required (POSIX atomic appends)
• Fast brain can read while slow brain writes
• System naturally converges on next read

Inspired by Cursor's self-driving codebase research.

License

MIT © Emotion Machine, Inc.