A Java AI agent platform that routes around dead LLM vendors

What it does

MateClaw is a multi-agent orchestration platform written in Java and Spring Boot. It lets you deploy persistent “digital employees”—AI agents with roles, memory, and tool access—across web, desktop, and chat channels like Slack or Telegram. Everything ships as a single JAR meant to run on your own server, with MySQL for production and an embedded H2 database for local use.

The interesting bit

The project treats LLM providers like unreliable microservices: a health tracker parks dead vendors in cooldown and routes requests to the next live model mid-conversation, so a provider outage doesn’t stop your team. That same operational mindset shows up in per-employee tool binding, approval-gated sensitive actions, and a runtime console where you can force-recycle a stuck agent.

Key highlights

• Multi-vendor failover with health tracking and cooldown for 14+ providers including OpenAI, Anthropic, DeepSeek, and Ollama
• LLM Wiki that digests raw files into linked, citation-backed pages and auto-injects them into agent prompts
• ReAct and Plan-and-Execute reasoning on a StateGraph runtime, with parallel delegation between agents
• Per-employee RBAC, approval flows for risky tool calls, and a full audit trail
• Eight IM channels, a web widget, desktop app, and web admin console from one deployment

Caveats

• Community is smaller than comparable personal-agent frameworks like OpenClaw or Hermes Agent, per the project’s own comparison
• Heavily Java-centric; the entire architecture assumes a Spring Boot operations model

Verdict

Grab this if you run a Java shop and need a self-hosted, team-aware agent platform with real governance and failover. Skip it if you want a lightweight personal CLI assistant or if Spring Boot dependencies feel like overkill for a side project.