Skip to content

Promote HomeClaw: Companion App, Plugins, Skills, Mix Mode, and Multi-Agent

A short guide to five strengths of HomeClaw: one companion app on all devices, extend with plugins, reuse OpenClaw skillsets, save cost with mix mode, and run multiple agents with no extra orchestration.

1. Companion App: One App, All Devices

The HomeClaw Companion is a Flutter-based client for Mac, Windows, iPhone, and Android. Same codebase, one install—chat and manage Core from any device.

  • Chat — Send messages, attach images and files; voice input and TTS (speak replies).
  • Manage Core — Edit core.yml and user.yml from the app (server, LLM, memory, tools, auth, etc.). No need to SSH or edit config files by hand.
  • Use it instead of or together with WebChat, CLI, Telegram, and other channels—all talk to the same Core and memory.

Takeaway: One app for desktop and mobile; chat, voice, and config from your phone or laptop. See Companion app.

2. Plugins: Extend HomeClaw Your Way

One plugin = one capability. Weather, news, email, browser automation, custom APIs—you add what you need.

Built-in plugins (Python)

  • Live in plugins/<Name>/: plugin.yaml, config.yml, plugin.py.
  • Core discovers them at startup. The LLM sees the list and calls route_to_plugin(plugin_id) when the user’s intent matches.
  • Examples: Weather, News, Mail.

External plugins (any language)

  • Write in any language: Node.js, Go, Java, Python, etc.
  • Run your plugin as an HTTP server with:
  • GET /health → 2xx
  • POST /run (or your path) → PluginRequest in, PluginResult out.
  • Register with Core: POST http://<core>:9000/api/plugins/register with id, description, health URL, and capabilities.
  • After that, Core routes to your server like built-in plugins. No need to write Python if you prefer another stack.

Takeaway: Add features without forking Core. Use Python for quick built-ins, or any language for external services.

3. Skills: Reuse OpenClaw Skillsets

HomeClaw supports OpenClaw-style skills: each skill is a folder under skills/ with a SKILL.md (name, description, workflow). The LLM sees the skill list and uses tools to accomplish workflows, or calls run_skill to run scripts under skill/scripts/.

You can reuse skillsets from OpenClaw: put the same skill folders (with SKILL.md and any scripts) into HomeClaw’s skills/ and they work. No rewrite—same format, same idea. HomeClaw loads them at startup and injects the skills block into the prompt so the model can choose and run them.

Takeaway: OpenClaw skills are compatible. Copy skill folders into skills/ and use them in HomeClaw as-is.

4. Mix Mode: Save Cost by Routing Smart

Problem: Cloud APIs are powerful but cost money. Local models are free but may be weaker on hard tasks.

Mix mode: For each user message, HomeClaw decides once: use the local model or the cloud model for the whole turn. Simple or private tasks → local. Search, complex reasoning, or “I’m not sure” → cloud. You get one model per turn, chosen automatically.

How it works

  1. Set in config/core.yml:
  2. main_llm_mode: mix
  3. main_llm_local: local_models/<your_local_id>
  4. main_llm_cloud: cloud_models/<your_cloud_id>
  5. A 3-layer router runs before tools and plugins, using only the user message:
  6. Layer 1 (heuristic): Keywords and rules (e.g. “hello”, “thanks” → local).
  7. Layer 2 (semantic): Similarity to example utterances (local vs cloud).
  8. Layer 3 (optional): Small classifier or perplexity probe.
  9. If no layer selects, default_route is used (local to save cost, or cloud for safety).

Cost control

  • default_route: local → when in doubt, stay local (fewer cloud calls).
  • Tune semantic threshold: higher = fewer “semantic” matches, more fallback to default or Layer 3.
  • Use usage_report (in chat or GET /api/reports/usage) to see how many requests went local vs cloud and tune rules.

Takeaway: Use local for cheap, simple, or private tasks; use cloud only when needed. One config, automatic per-request routing, and visibility into usage.

5. Multi-Agent: Run as Many Agents as You Need

One HomeClaw instance = one agent (one memory, one set of tools and plugins). To get multiple agents, run multiple instances.

How to do it

  • Start each instance with a different port and (if you want) a different config.
  • Example: Agent A on port 9000, Agent B on port 9001. Different core.yml (or same file, different server.port).
  • No central orchestrator: each instance is independent. Point your clients (Companion app, WebChat, Telegram, etc.) at the right port.

When it helps

  • Separate roles: e.g. one agent for home automation, one for work notes.
  • Isolation: different users or projects on different instances.
  • Scale: add more machines or more processes; each process is one agent.

Takeaway: Multi-agent = multi-instance. Run more processes (different ports/configs); no extra orchestration layer.

Summary

Topic In one sentence
Companion app Flutter app for Mac, Windows, iPhone, Android: chat, voice, attachments, and edit core.yml/user.yml from the app.
Plugins Add capabilities with built-in (Python) or external (any language) plugins; register once, LLM routes to them.
Skills Reuse OpenClaw skillsets: same SKILL.md format in skills/; copy skill folders over and they work.
Mix mode Per-request routing: local for simple/private, cloud for hard tasks; 3-layer router + usage reports to control cost.
Multi-agent One instance = one agent; run multiple instances (different ports/configs) for multiple agents, no orchestrator.

For details, see Companion app, Plugins, Tools (skills), Mix mode and reports, and Introducing HomeClaw.