OpenClaw Voice Control

Local macOS voice-control integration for OpenClaw. Use when setting up, deploying, troubleshooting, or operating wakeword-triggered voice access to a local...

Security Scan

The skill's requirements and instructions are coherent with a macOS local voice-control integration for OpenClaw, but it will install and run code from the referenced GitHub repo and may enable a background auto-start — inspect and confirm before proceeding.

Purpose & Capability

Name, description, required binaries (python3, git, launchctl), required env vars (OPENCLAW_BASE_URL, OPENCLAW_TOKEN, SENSEVOICE_MODEL_PATH, SENSEVOICE_VAD_MODEL_PATH) and launchd usage align with a local macOS voice-control integration for OpenClaw.

Instruction Scope

SKILL.md directs the agent to clone the GitHub repo into the skill workspace, create/activate a venv, pip install -e ., download local models, run the voice service and an overlay, and optionally deploy a launchd auto-start. These steps are within scope but involve running repository code and network downloads; the skill does explicitly require user confirmation before system-changing steps.

Install Mechanism

There is no automated install spec in the registry (instruction-only). The recommended flow uses git fetch and pip install -e . from the upstream GitHub repo and model downloads (modelscope/funasr). This is expected for this kind of local integration, but pip install will execute the repo's installation code so the repo contents should be reviewed prior to installation.

Credentials

Requested environment variables are appropriate for the stated purpose: a base URL and token to contact the local OpenClaw agent and local model path variables for ASR/VAD. No unrelated credentials or broad cloud creds are requested. Note that OPENCLAW_TOKEN is a primary credential that grants whatever access the OpenClaw service allows, so use least-privilege tokens.

Persistence & Privilege

The skill does not force permanent inclusion (always:false). It explicitly offers optional background resident behavior and launchd auto-start via a deploy script; this is reasonable for a persistent voice entrypoint but increases attack surface and should only be enabled after inspecting deploy_macos.sh and confirming intent.

Assessment

This skill appears to do what it says, but it will fetch and install code from the linked GitHub repository and download local ML models. Before installing: (1) review the repository (especially setup/pyproject, setup scripts, and scripts/deploy_macos.sh) to confirm there is no unexpected behavior, (2) do not supply a highly-privileged OPENCLAW_TOKEN — create a least-privilege token for voice access, (3) confirm any steps that modify system state (launchd auto-start), and (4) prefer running the initial validation in an isolated user account or VM if you want to limit risk. If you are not comfortable reviewing the repo, do not run pip install -e . or enable auto-start.

OpenClaw Voice Control is a local macOS voice-control integration for OpenClaw.

Repository source:

• GitHub: CarrotYuan/openclaw-voice-control

It provides:

• wakeword activation
• local microphone capture
• local ASR with FunASR / SenseVoice
• forwarding recognized text to a local OpenClaw agent
• macOS TTS playback
• optional overlay UI
• launchd-based background resident behavior
• auto-start after user login

Platform

• macOS only

Safety Warning

This skill adds a voice entrypoint, and that entrypoint is not identity-bound.

That means:

• anyone near the microphone may be able to trigger it
• any capability exposed through the connected OpenClaw agent may become reachable by voice

Recommended safety practice:

• add explicit safety constraints to the connected agent prompt
• require confirmation for high-risk actions
• avoid broad autonomous permissions for the voice-facing agent
• use least privilege for tools and connected systems

Main Path

Treat this as a local deployment skill, not as a prompt-only helper.

When this skill is installed into OpenClaw:

• work inside the current installed skill directory for the current conversation agent
• use the declared GitHub repository as the source of truth
• do not silently switch to another local clone or an already-prepared environment
• do not continue until the workspace contains the real repository files, not just SKILL.md

The main installation guide is the repository README.md. README.zh-CN.md is the Chinese companion guide.

Use this as the standard install path:

1. sync the full repository into the current installed skill workspace
2. create and activate .venv
3. run pip install -e .
4. download or populate the SenseVoice model directory
5. download or populate the VAD model directory
6. copy .env.example to .env
7. fill the required values in .env
8. use the default openWakeWord route
9. run direct-run validation with both the voice service and overlay process
10. ask whether the user wants background resident behavior and auto-start
11. if yes, run ./scripts/deploy_macos.sh
12. if no, stop after direct-run validation

Before running any system-changing step in that path, explicitly tell the user what you are about to do and get confirmation for:

• fetching and checking out the repository into the skill workspace
• running pip install -e . from the repository
• downloading large local models or wakeword assets
• enabling background resident behavior or launchd auto-start

Minimum command path:

# from the current conversation agent's installed skill directory
git init
git remote add origin https://github.com/CarrotYuan/openclaw-voice-control.git
git fetch --depth 1 origin main
git checkout -B main FETCH_HEAD
# do not continue until the workspace contains scripts/, src/, config/,
# launchagents/, and README.md
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
./.venv/bin/modelscope download --model iic/SenseVoiceSmall --local_dir models/SenseVoiceSmall
./.venv/bin/python - <<'PY'
from funasr import AutoModel
AutoModel(model='fsmn-vad', disable_update=True)
PY
cp .env.example .env
# terminal 1, from the current installed skill workspace
python -m openclaw_voice_control --config config/default.yaml --env-file .env
# terminal 2, from the same workspace
python -m openclaw_voice_control.overlay_app --config config/default.yaml --env-file .env

Direct-run validation is not complete unless both commands above are running at the same time from the same installed skill workspace.

Before any next step after direct-run validation, stop that test first.

This includes:

• background resident deployment
• auto-start validation
• starting another direct-run test

If an old direct-run service and overlay are left running, two active voice runtimes can respond to the same wakeword and produce duplicate replies.

What Must Exist Before Setup

Before setup begins, make sure these prerequisites exist or can be provided:

• Python 3.11 or newer
• a running local OpenClaw service
• OPENCLAW_TOKEN
• macOS microphone permission
• a way to download the local SenseVoice and VAD model directories

For the default route, the remaining setup can usually be handled by the AI or operator:

• copy .env.example to .env
• download the SenseVoice model directory
• download or populate the VAD model directory
• let openWakeWord download the selected pretrained wakeword model on first run

Recommended Defaults

The default public route is:

• WAKEWORD_PROVIDER=openwakeword
• OPENWAKEWORD_MODEL_NAME=hey jarvis
• OPENCLAW_AGENT_ID=main
• OPENCLAW_MODEL=openclaw:main
• OPENCLAW_USER=openclaw-voice-control

Prefer that route unless the user explicitly asks for something else.

Default Route Variables

These are the main values for the default route:

• required in .env
  • OPENCLAW_BASE_URL
  • OPENCLAW_TOKEN
  • SENSEVOICE_MODEL_PATH
  • SENSEVOICE_VAD_MODEL_PATH
• usually left at their documented defaults unless the user wants customization
  • WAKEWORD_PROVIDER=openwakeword
  • OPENWAKEWORD_MODEL_NAME=hey jarvis
  • OPENCLAW_AGENT_ID=main
  • OPENCLAW_MODEL=openclaw:main
  • OPENCLAW_USER=openclaw-voice-control

If the user changes the macOS TTS voice, first make sure that voice has been downloaded in:

• System Settings
• Accessibility
• Spoken Content
• the i button next to System Voice
• Language
• Voice

Optional Porcupine Route

Picovoice / Porcupine is an optional fallback route, not the default path.

Only switch to it when the user explicitly wants to use a local .ppn wakeword model.

If that route is chosen, set:

• WAKEWORD_PROVIDER=porcupine
• PICOVOICE_ACCESS_KEY
• WAKEWORD_FILE

Only ask for those Porcupine-specific values when the user explicitly chooses that route.

Switching openWakeWord Models

To switch the default openWakeWord wakeword, change:

• OPENWAKEWORD_MODEL_NAME

Common official pretrained examples include:

• hey jarvis
• hey mycroft
• hey rhasspy
• alexa

Those pretrained wakeword models are downloaded automatically on first use.

The code path already supports changing OPENWAKEWORD_MODEL_NAME, but only the default hey jarvis route has been smoke-tested in this repository so far.

Execution Rules

When using this skill, follow these rules:

1. Use the declared GitHub repository first.

   • Start from https://github.com/CarrotYuan/openclaw-voice-control.git
   • Do not substitute another repository just because it looks similar.

2. Keep deployment work in the current installed skill workspace.

   • Do not silently switch to another local clone.
   • If GitHub access fails and a local clone exists, ask before using it.

3. Do not silently reuse an old environment.

   • Do not assume an existing .venv, local model cache, previous .env, or private runtime should be reused.
   • If reusing cached assets might save time, explain that and ask first.

4. Do not invent missing values.

   • When writing .env, use the exact variable names required by the project, especially OPENCLAW_TOKEN.
   • If required values or local assets are missing, stop and point to the documented source of truth.

5. Handle secrets conservatively.

   • Do not print real tokens or keys back into the conversation unless the user explicitly asks to see them.
   • For this project, use the OpenClaw token from ~/.openclaw/openclaw.json, specifically the gateway configuration.
   • Do not use ~/.openclaw/identity/device-auth.json as the token source for this project.

6. Ask before any system-changing action.

   • Do not assume you should fetch the repository, run pip install -e ., download models, or enable launchd behavior without user approval.
   • Explain the action first, then continue only after the user confirms.

7. Ask before enabling background resident behavior.

   • Foreground validation comes first.
   • Direct-run validation means starting both python -m openclaw_voice_control --config config/default.yaml --env-file .env and python -m openclaw_voice_control.overlay_app --config config/default.yaml --env-file .env from the same installed skill workspace.
   • Only run ./scripts/deploy_macos.sh when the user explicitly wants background resident behavior or auto-start.

Daily Maintenance

Primary scripts:

• deploy background runtime: ./scripts/deploy_macos.sh
• restart installed background services: ./scripts/restart_service.sh
• uninstall installed background services: ./scripts/uninstall_macos.sh
• inspect local installation and environment issues: ./scripts/doctor.sh

Double-click .command wrappers are also available in scripts/ for macOS users who prefer Finder-based execution.

Shutdown Intents

Treat shutdown requests as one of these two user intents:

• temporarily disable voice functionality
  • stop the running direct-run process, or stop the deployed background runtime
  • do not delete the skill folder
• delete the skill completely
  • remove the skill folder itself
  • only do this when the user explicitly asks for full removal

If the user says something ambiguous like "turn it off" or "stop voice", ask one short clarification question before acting.

Background Architecture

The canonical background startup path is:

• launchd -> host app -> shell script -> python

./scripts/deploy_macos.sh builds the required host apps automatically.

Configuration Surface

Default-route required values:

• OPENCLAW_BASE_URL
• OPENCLAW_TOKEN
• SENSEVOICE_MODEL_PATH
• SENSEVOICE_VAD_MODEL_PATH

Default-route configurable values:

• OPENCLAW_AGENT_ID
• OPENCLAW_MODEL
• OPENCLAW_USER
• WAKEWORD_PROVIDER
• OPENWAKEWORD_MODEL_NAME
• OPENWAKEWORD_MODEL_PATH

If the user explicitly switches to the optional Porcupine route, also configure:

• PICOVOICE_ACCESS_KEY
• WAKEWORD_FILE

Interpreter override variables still exist:

• VOICE_CONTROL_PYTHON_BIN
• VOICE_CONTROL_OVERLAY_PYTHON_BIN

They are troubleshooting knobs only, not the main deployment model.

OPENWAKEWORD_THRESHOLD still exists as a tuning variable, but it is a troubleshooting adjustment rather than a first-run requirement.

As a rule, machine-specific secrets and paths belong in .env, while wakeword timing and threshold tuning should normally be adjusted in config/default.yaml.

Where To Get Missing Requirements

After syncing the repository into the current installed skill workspace, read these sections in README.md:

• What Must Exist Before Setup
• Required Variables
• Where To Get Each Requirement

Practical source notes:

• OPENCLAW_BASE_URL: use the full OpenClaw chat completions endpoint, not only the host and port root. For the default local setup, use http://127.0.0.1:18789/v1/chat/completions
• OPENCLAW_TOKEN: obtain it from the local OpenClaw gateway configuration in ~/.openclaw/openclaw.json, under gateway
• default wakeword route: use openWakeWord with the built-in English hey jarvis model
• optional Porcupine route: obtain PICOVOICE_ACCESS_KEY and the local .ppn file from Picovoice
• if GitHub clone fails, report that first rather than switching to an unrelated local directory

Related Files

• README.md in the cloned repository
• README.zh-CN.md in the cloned repository
• docs/macos-install.md in the cloned repository