MateClaw — Summary

MateClaw is a Spring Boot 3.5 / Java 21 multi-agent harness platform designed for organizational deployment — not single-user hacking. Built by the matevip team, it ships as a single JAR with an embedded Vue 3 admin SPA, Electron desktop binaries with bundled JRE 21, an embeddable webchat widget, and eight IM channel adapters (DingTalk, Feishu, WeChat Work, WeChat, Telegram, Discord, QQ, Slack). The architecture centers on Digital Employees (ReAct + Plan-and-Execute agents backed by a Spring AI Alibaba StateGraph runtime) with RBAC per agent, per model, and per tool; an approval gate for sensitive tool calls; a full audit trail; and a runtime console for live visibility into all running agents. MateClaw's multi-vendor failover (chain with health tracker and cooldown window) is explicitly designed to handle API outages: when the primary provider returns 401, the next healthy provider picks up mid-sentence. Business orchestration in 1.3.0+ adds Workflows (7-step-mode linear DSL), Triggers (6 pattern types: cron/webhook/channel_message/agent_lifecycle/content_match/workflow_completion), and a LLM Wiki with citation-linked pages and a Transformations engine.

differs_from_seeds: MateClaw is closest to claude-flow in scope (MCP integration, multi-agent, rich memory/knowledge systems) but runs on the JVM with enterprise-grade access control absent from all seeds. Unlike claude-flow's file-based tasks.json or BMAD's persona-md files, MateClaw defines agents as Spring Boot service beans with role/goal/backstory and manages them via a web admin console. The Workflow DSL, 8-channel IM adapters, RBAC, per-agent model binding, and "always-on" philosophy (one JAR deployed as a server, not a CLI run on demand) distinguish it from every seed in the corpus.

MateClaw — Overview

Origin

MateClaw is developed by matevip (GitHub: matevip/mateclaw). The README explicitly contrasts it with single-user tools:

"Other personal AI agents are built for one person. MateClaw is the one your IT department can actually sign off on."

The project emerged as the enterprise-ready counterpart to OpenClaw, targeting organizations (Spring Boot shops) that need RBAC, approval flows, audit trails, and multi-instance deployment.

Philosophy

Three core differentiation claims (from README):

1. Multi-vendor failover:

"Primary key expired. Vendor returns 401. Network blip. Quota drained. Other tools hand you a red error card. MateClaw routes to the next healthy provider — DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio, MLX, 14+ in total — and the user sees the reply finish."

2. Knowledge that links itself:

"MateClaw's LLM Wiki digests [raw materials] into structured pages, builds [[links]] between them, and remembers where every sentence came from. Click a citation, see the exact source chunk."

3. One product, five surfaces:

"Web Console, Desktop (Electron), Webchat Widget, IM Channels, Plugin SDK. Same brain. Same memory. Same tools. Different doors."

The "Always-On" Position

MateClaw is a server — always running, deployed as infrastructure. This is the central "always-on" boundary in this batch: MateClaw doesn't start when you run a command; it runs continuously and handles incoming IM messages, webhook triggers, cron jobs, and workflow events regardless of whether a user is at their keyboard.

The AI-as-Infrastructure Framing

"On March 2, 2026, Claude went dark for 4 hours across API, web, and mobile. Three weeks later, another 5 hours. Every company that bet their AI strategy on a single vendor spent those outages staring at red error cards... MateClaw is that layer — built the Spring Boot way."

Five Career Templates

Digital Employee career templates shipped:

1. Product Researcher
2. Customer Support
3. Knowledge Curator
4. Data Analyst
5. Executive Assistant

MateClaw — Architecture

Distribution

• Type: Spring Boot JAR + Electron desktop binary
• License: Apache-2.0
• Backend runtime: Java 21+
• Frontend: Vue 3 + TypeScript (built into server JAR)
• Database: MySQL (production), H2 (dev — auto-selected)
• Schema management: Flyway (auto-heals on upgrade)

Install Options

# Maven (backend)
cd mateclaw-server && mvn spring-boot:run  # http://localhost:18088

# Frontend (dev)
cd mateclaw-ui && pnpm install && pnpm dev  # http://localhost:5173

# Docker
cp .env.example .env && docker compose up -d  # http://localhost:18080

# Desktop
# Download Electron binary with bundled JRE 21 from GitHub Releases

Project Structure

mateclaw/
├── mateclaw-server/        # Spring Boot 3.5 backend
│   └── src/main/java/vip/mate/
│       ├── agent/          # AgentService, AgentState, BaseAgent, StateGraph
│       ├── skill/          # Skill installer, manifest, lessons, runtime
│       ├── workflow/       # Workflow DSL (7 step modes)
│       ├── trigger/        # 6 trigger pattern types
│       ├── channel/        # 8 IM channel adapters
│       ├── wiki/           # LLM Wiki with citations + Transformations
│       ├── memory/         # Memory lifecycle (extraction, consolidation, dreaming)
│       ├── tool/           # Tool registry with RBAC + approval
│       ├── audit/          # Full audit trail
│       └── mcp/            # MCP client (stdio, SSE, Streamable HTTP)
├── mateclaw-ui/            # Vue 3 admin SPA
├── mateclaw-webchat/       # Embeddable widget (UMD/ES)
├── mateclaw-plugin-api/    # Java SDK for third-party plugins
└── mateclaw-plugin-sample/ # Reference plugin

Tech Stack

Provider Failover

Health tracker monitors 14+ LLM providers. Bad providers enter a cooldown window. Routing order configurable via drag-and-drop in Settings → Models.

MateClaw — Components

Digital Employees (Agents)

Each "Digital Employee" has:

• Role — job title/function
• Goal — what they optimize for
• Backstory — personality and context
• Avatar — pixel-art image + color
• Career template — one of 5 built-in types

Reasoning modes: ReAct (iterative tool-calling) or Plan-and-Execute (decompose → execute steps).

Skills System

• Format: SKILL.md manifest + prompt + tool list + LESSONS.md
• LESSONS.md: Grows smarter with use — records learnings from each skill invocation
• 8 starter templates included
• 5-step creation wizard in the UI
• Pre-flight checks: validation before install (missing deps flagged)
• Per-employee binding (1.3.0+): tool installed for one employee doesn't bleed to others

MCP Integration

• Protocols: stdio, SSE, Streamable HTTP
• Per-employee MCP binding (1.3.0+)

ACP (Agent Communication Protocol)

• Brings Claude Code, Codex, etc. as "employees" via ACP adapter
• Auto-bridged to skill cards with wrapper tools

IM Channel Adapters (8)

DingTalk, Feishu, WeChat Work, WeChat, Telegram, Discord, QQ, Slack

Workflow DSL (7 step modes)

JSON-first authoring with Monaco + schema validation, or natural-language → draft generation.

Triggers (6 types)

cron / webhook / channel_message / agent_lifecycle / content_match / workflow_completion

Default-on governance: dedup, per-trigger rate limit, bot-self filter, recursion guard, fail-closed unknown patterns.

LLM Wiki + Transformations

• Raw materials (PDFs, markdown, scraped pages) digested into linked wiki pages with citations
• Hot cache auto-injects relevant pages into system prompts
• Transformations engine (1.3.0+): user-authored templates run against wiki content; cross-material map-reduce; JSON output mode; per-template model picker

Memory Lifecycle

• Post-conversation extraction
• Scheduled consolidation
• Dreaming workflows
• write_memory workflow step
• Memory files: AGENTS.md, SOUL.md, PROFILE.md, MEMORY.md, daily notes

Admin Runtime Console

Real-time visibility: who's running, what step, token counts, force-recycle in one click.

Document Generation Tools

DocxRenderTool, XlsxRenderTool, PptxRenderTool, PdfRenderTool — render Markdown to Office files inside the JVM (no subprocess, no Office install).

Tool Guard

RBAC + approval flow + path protection. Per-employee tool binding.

Multimodal

Text-to-speech, speech-to-text, image, music, video, 3D. Sidecar routing: text-only main model + image attachment → vision sidecar describes → main model answers.

MateClaw — Uniqueness

differs_from_seeds

MateClaw is closest to claude-flow in scope (multi-agent, MCP, rich memory) but operates on a completely different stack and boundary. claude-flow is a developer tool (npm package, SQLite, file-based) used from the terminal; MateClaw is an organizational server (Spring Boot JAR, MySQL, Flyway) that runs continuously and handles incoming events from IM channels. The "always-on" execution mode (background-daemon waiting for IM messages, webhooks, and cron jobs) is architecturally absent from all 11 seeds — every seed is a tool you invoke, whereas MateClaw is infrastructure you deploy. The 14-provider failover chain with health tracker and cooldown is the most production-hardened availability story in this batch. The RBAC+approval-gate+audit-trail combination, the 8 IM channel adapters, the Workflow DSL with await_approval steps, and the Java/Spring Boot runtime make MateClaw the most enterprise-aligned framework in the entire corpus. The LESSONS.md adaptive skill evolution is also not present in any seed.

Positioning

MateClaw positions itself as the AI platform for "the IT department" — the answer to "we need RBAC, approval workflows, multi-user isolation, a full audit trail, and the ability to survive a provider outage" in the same conversation.

Observable Failure Modes

• Setup complexity: Java 21 + Spring Boot + MySQL + optionally Docker — more infrastructure than any seed
• Small community: 512 stars vs OpenClaw's larger community (explicitly acknowledged in README)
• Chinese-first origin: README, architecture docs, and issue tracker may be Chinese-language primary
• Enterprise features require deployment: This is not a laptop tool — you need a server, a database, and network access for full functionality
• No CLI dev workflow: Engineers accustomed to claude or codex commands will find the admin UI + IM-channel-based workflow unfamiliar
• SKILL.md licenses: Proprietary licenses on document skills (inherited from similar projects)