An enterprise-grade multi-agent AI assistant platform for teams, building a "second brain" through multi-model failover, MCP protocol, and multi-channel support.
Positioning & Use Cases#
MateClaw is positioned as an upper orchestration and scheduling layer. It does not involve training or fine-tuning underlying LLMs, but focuses on multi-model routing, RAG knowledge digestion, and unified multi-channel distribution. Suitable for enterprise AI middleware governance, multi-model failover scheduling, team private knowledge base Q&A, unified multi-channel intelligent customer service, and private deployment environments with strict data security requirements.
Core Capabilities#
Model Scheduling & Failover#
Integrates 14+ model providers including DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio, and MLX. Built-in Provider Health Tracker continuously monitors provider health; upon detecting failures, a cooldown window is triggered and subsequent requests are seamlessly routed to backup nodes, ensuring business continuity.
Agent Runtime#
Based on the StateGraph engine, supporting ReAct (iterative reasoning) and Plan-and-Execute (complex multi-step workflow) modes. Features dynamic context pruning, intelligent truncation, and expired stream cleanup for long conversation stability.
Knowledge Base (LLM Wiki)#
Supports uploading PDFs, Markdown, and web pages, automatically digesting them into structured pages with [[links]] bidirectional link relationships. Q&A results are assembled from relevant chunks, with every sentence traceable to the original text via clickable citations.
Tools, Skills & MCP Protocol#
Built-in tools: web search, file operations, memory management. Native MCP (Model Context Protocol) support via stdio / SSE / Streamable HTTP transports. Integrates ClawHub skill marketplace (SKILL.md skill packs). Tool Guard layer provides tool-level RBAC, approval flows, and path protection.
Memory System#
Maintains Workspace memory files (AGENTS.md, SOUL.md, PROFILE.md, MEMORY.md, daily notes). Implements a full lifecycle: post-conversation extraction → scheduled integration → Dreaming workflow.
Multimodal Support#
TTS, STT, image, music, and video are all first-class citizens.
Multi-Channel Access#
Covers Web Console (full management backend), Electron desktop app (bundled JRE 21, no Java installation needed), Webchat Widget (<script> embed), and 8 IM channels: DingTalk, Feishu/Lark, WeCom, WeChat, Telegram, Discord, QQ, Slack.
Enterprise Features#
RBAC + JWT authentication, full audit logs, Flyway database auto-migration, Spring Boot Actuator monitoring, per-channel error isolation, single JAR deployment.
Architecture & Design Philosophy#
Module Structure#
Monorepo structure:
mateclaw-server: Spring Boot 3.5 backend core (Spring AI Alibaba 1.1, StateGraph runtime)mateclaw-ui: Vue 3 + TypeScript admin SPA (bundled into server JAR)mateclaw-webchat: Embeddable chat component library (UMD / ES bundles)mateclaw-plugin-api/mateclaw-plugin-sample: Third-party plugin development framework
Key Design Principles#
- "Whole widget" philosophy: Reasoning, knowledge, memory, tools, and channels are built as an integrated whole, not loosely assembled.
- Per-channel Error Isolation: A single IM channel failure does not affect other channels.
- Memory Lifecycle: Automated workflows to consolidate structured memories.
Installation#
Local Development#
cd mateclaw-server
mvn spring-boot:run # http://localhost:18088
cd mateclaw-ui
pnpm install && pnpm dev # http://localhost:5173
# Default login: admin / admin123
Docker#
cp .env.example .env
docker compose up -d # http://localhost:18080
Desktop App#
Download from GitHub Releases. Bundled JRE 21, double-click to run.
Unconfirmed Information#
- ClawHub skill marketplace independent URL not publicly available
- Live Demo specific URL not displayed
- Hugging Face model/dataset page not mentioned
- Associated academic paper not found