HomeClaw Adds Plugins, Mix Mode & Multi-Agent Support

HomeClaw, an extensible home automation and AI assistant framework, has released a major update introducing five key architectural enhancements: a plugin system supporting multiple languages, OpenClaw skill compatibility, intelligent cost-saving model routing, multi-agent deployment patterns, and a cross-platform companion app. These additions position HomeClaw as a flexible platform for developers building AI-powered automation systems.

Plugin Architecture: Language-Agnostic Extensions

HomeClaw's new plugin system enables developers to extend functionality without modifying core code. Built-in Python plugins live in the plugins/ directory with a standard structure: plugin.yaml, config.yml, and plugin.py. The LLM automatically discovers plugins at startup and routes user intents via route_to_plugin(plugin_id).

External plugins can be written in any language—Node.js, Go, Java, or Python—by running an HTTP server that implements two endpoints: GET /health and POST /run. Developers register external plugins with Core via POST http://host:9000/api/plugins/register, providing an ID, description, health URL, and capabilities. Core then routes requests to external services identically to built-in plugins, eliminating the need to work in Python if another stack is preferred.

This architecture allows teams to build capabilities in their language of choice while maintaining a unified interface. Weather services, email automation, browser control, and custom API integrations can all be added as independent plugins without forking the core repository.

OpenClaw Skill Compatibility

HomeClaw now supports OpenClaw-style skills, enabling direct reuse of existing skillsets. Each skill is a folder under config/skills/ containing a SKILL.md file that defines the skill name, description, and workflow. The LLM reads the skill list at startup and either uses built-in tools to accomplish workflows or calls run_skill to execute scripts in skill/scripts/.

Developers can copy skill folders from OpenClaw projects directly into HomeClaw's config/skills/ directory with no modifications required. The same format and structure work across both platforms, allowing teams to leverage existing automation workflows immediately. This compatibility reduces migration friction and enables knowledge sharing across the Antfarm: Multi-Agent Workflow Orchestration for OpenClaw ecosystem.

Mix Mode: Cost-Optimized Model Routing

The Mix Mode feature addresses a common challenge in AI application development: balancing the power of cloud APIs against the cost savings of local models. For each user message, HomeClaw automatically decides whether to route the request to a local model or a cloud API based on task complexity and privacy requirements.

The routing system employs a three-layer decision process before invoking tools or plugins. Layer 1 uses heuristic rules and keywords to catch simple interactions like greetings or confirmations. Layer 2 applies semantic similarity matching against example utterances to classify intent. Layer 3 optionally uses a small classifier or perplexity probe for edge cases. If no layer makes a confident decision, the system falls back to a configurable default route—typically local to minimize costs.

Configuration is straightforward in config/core.yml: set main_llm_mode: mix, specify main_llm_local and main_llm_cloud endpoints, and choose a default_route. Developers can tune the semantic threshold to control routing sensitivity and use the usage_report endpoint to monitor local versus cloud request distribution. This visibility enables iterative optimization of routing rules to balance cost and quality.

Multi-Agent Deployment Pattern

HomeClaw implements multi-agent architectures through a simple pattern: each agent is an independent instance with its own memory, tools, and plugins. Developers run multiple HomeClaw processes on different ports, each with an optional separate configuration file. For example, Agent A might run on port 9000 for home automation while Agent B runs on port 9001 for work-related tasks.

This approach provides clear isolation between agents with different roles, users, or projects. There is no central orchestrator—each instance is fully independent, and clients like the Companion app, WebChat, or Telegram bots connect to the appropriate port. The pattern scales horizontally by adding more processes across machines without architectural complexity.

Cross-Platform Companion App

The HomeClaw Companion app, built with Flutter, delivers a unified experience across Mac, Windows, iPhone, and Android from a single codebase. The app provides a full-featured chat interface with image and file attachment support, voice input, and text-to-speech for spoken replies.

A key differentiator is the built-in configuration management interface. Users can edit core.yml and user.yml directly from the app, adjusting server settings, LLM providers, memory configuration, tool permissions, and authentication without SSH access or manual file editing. The companion app works alongside other interfaces like WebChat, CLI, and Telegram—all channels share the same Core instance and memory state.

Implications for AI Developers

These updates reflect a broader trend toward modular, cost-conscious AI architectures. The plugin system's language independence lowers barriers to contribution and allows teams to integrate existing services without rewriting them. OpenClaw skill compatibility creates interoperability across related projects, building network effects in the open-source home-automation space.

Mix Mode's routing approach addresses a practical concern for production deployments: cloud API costs can quickly become prohibitive for high-volume applications. By automatically selecting the appropriate model tier based on task characteristics, developers can maintain quality on complex requests while reducing expenses on routine interactions. The built-in usage reporting enables data-driven optimization of routing rules.

The multi-instance pattern for multi-agent systems favors operational simplicity over architectural complexity. Rather than building orchestration layers, developers can leverage existing process management and load balancing tools. This aligns with microservices principles and makes HomeClaw deployments easier to monitor, debug, and scale.

Source: Medium article by Allenpengrocks, published February 2026. Project repository available at github.com/allenpeng0705/HomeClaw.