MateClaw
A self-hosted second-brain digital-employee platform: multi-provider failover, LLM Wiki knowledge digestion, skills, memory and multi-channel delivery — shipped as one Spring Boot JAR your IT department can sign off on.
README claims a fairly complete security narrative — Tool Guard (RBAC+approval+path protection), approval gates, audit trail, AES-GCM secrets, HMAC-SHA-256 webhook signing, explicit idempotency warnings for external side effects — but the sample contains only README, LICENSE, and two test files; no security implementation is visible. Least privilege, sensitive-data handling, and dependency security are assertions only, so conservative deductions apply. Approval gates and external-effects honesty (no exactly-once promise) are well described at the documentation level, earning middle scores. Rollback is claimed (skill snapshots/restore, DSH disable) but unverified. Attribution is limited to the LICENSE copyright line.
Internally self-consistent: version numbers, dates, and module layout agree; failover, health tracking, per-channel error isolation, and SSE reconnect IDs are described with concrete mechanisms. All are claims without code corroboration, and error-message quality is only indirectly evidenced ('Chinese error hints'), hence 2s.
Audience (teams/Java shops) and scenarios (web, desktop, widget, 8 IM channels, plugins) are clearly laid out; capability boundaries are stated repeatedly (no exactly-once for external side effects, DSH requires install/verification); environment fit (H2/Postgres/MySQL/Kingbase, bundled JRE 21) is thorough. Trigger precision (six pattern types, recursion guard, fail-closed) is described but has no implementation evidence in the sample.
The full Apache-2.0 text is present and matches the badge — full marks. Install notes cover Maven/npm, Docker, and desktop with default credentials. Project structure, naming, and a dated roadmap (1.8→2.2) are clear. However there is no standalone FAQ or known-limitations section, and no maintainer/contribution/update-path information, hence deductions.
Output usability (document tools, citation drawer, content calendar, runtime console) is concretely described; marginal value is positioned against named competitors; the cost model ($0, self-hosted, your keys) is explicit. But none of the claimed effectiveness is execution-verified in this static review, so 2s.
The README is assertion-dense (~78% load improvement, feature lists, competitor table) but relies on external doc links; no corresponding implementation is present in the sample to trace, and cross-source corroboration rests on just two small test files. Facts vs. inference are separable (README reads as claims), earning 2; claim traceability and corroboration are deducted.
- This review rests on README, LICENSE, and two small test files only; all security and reliability claims (RBAC, approvals, encryption, failover) are unverified against code — low-confidence static assessment.
- Default login admin/admin123 must be changed immediately after first startup, especially in the Docker deployment path.
- External side effects (publish, send, payments) are explicitly not promised exactly-once; deployers must add provider idempotency or manual approval.
- Publisher identity is unverified; supply-chain and maintenance-continuity risk must be assessed independently.
What does this agent do, and when should you use it?
MateClaw (github.com/mateaix/mateclaw) is a pluggable Agent Runtime built on Spring Boot 3.5 and Spring AI Alibaba, positioned as a team-oriented digital-employee platform. It ships a native StateGraph runtime (ReAct, Plan-and-Execute, persistent Goals, Team Runs) alongside a managed external DeepSeek Harness (DSH) JSON-RPC runtime, with both paths converging on the same conversations, workspace boundaries, Tool Guard, event projection, and lifecycle governance. On the knowledge side, the LLM Wiki digests uploaded PDFs and markdown into linked pages with traceable citations, complemented by AGENTS.md/SOUL.md/MEMORY.md workspace memory and Dream consolidation workflows. Capability extension covers SKILL.md packages, MCP (stdio/SSE/HTTP with per-employee binding), and an ACP bridge that brings Claude Code and Codex in as governed employees. One deployment serves five surfaces: the web admin console, an Electron desktop app with bundled JRE 21, an embeddable Webchat widget, eight IM channels (DingTalk, Feishu, WeChat Work, Telegram, etc.), and a Java plugin SDK. Enterprise features include RBAC + JWT, approval gates, audit trails, a runtime console, and multi-provider failover.
Users create digital employees (role/goal/backstory/avatar) in the web console and bind models, skills, MCP tools, and knowledge bases to them; each turn is dispatched via the AgentRuntimeProvider contract to either the native StateGraph (ReAct, Plan-and-Execute, Goal checklists, Team Run task DAG) or the DSH child process, streaming normalized events (thinking/tool calls/answers/usage) back to the UI. When a request fails, it retries down a priority chain of healthy providers (DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, etc.) with a health tracker cooling down bad vendors. Uploaded documents are digested into interlinked LLM Wiki pages with citations; the Workflow engine composes multi-employee flows using seven step modes (sequential/fan_out/collect/conditional/await_approval/dispatch_channel/write_memory), and Triggers support six pattern types including cron, webhook, and channel_message. Multimodal output spans TTS/STT, image, music, video, and 3D, while DocxRenderTool/XlsxRenderTool/PptxRenderTool/PdfRenderTool render Office/PDF files inside the JVM. Durable Goals persist checkpoints so long tasks resume after a backend restart.
- Java-stack teams needing a self-hosted, auditable, approval-gated multi-user AI employee platform their IT department will approve
- Support/product teams wiring DingTalk, Feishu, WeChat Work, Telegram and more into one memory-equipped assistant that shares a brain across channels
- Content operators using Content Studio to turn one sentence into WeChat Official Account articles or Xiaohongshu 3:4 card notes, with de-AI scoring and compliance scanning into draft boxes
- Users running hours-long tasks (bulk file production, ongoing research) who need persistent Goals to resume from checkpoints after restarts
- Engineering teams onboarding coding agents like Claude Code / Codex as governed employees with shared permissions, approvals, and memory
- Ops staff automating multi-employee collaboration via Workflows fired by cron or webhook triggers
What are this agent's strengths and limitations?
- Provider chain + health tracker + cooldown means one vendor outage doesn't take the product down — you just drag priorities in settings
- A real pluggable runtime contract (AgentRuntimeProvider): native StateGraph and external DSH share sessions, approvals, tools, and audit, so employees aren't welded to one reasoning loop
- Five delivery surfaces (web admin, desktop, widget, 8 IM channels, Java plugin SDK) share the same brain and memory
- Full enterprise governance: RBAC + approval gates + audit trail + distributed cron lock + runtime console force-recycle
- LLM Wiki provides page-level wikilinks and a citation drawer for verifying answer-to-source-chunk references
- Stack is bound to Java 21 + Spring Boot 3.5 + Vue 3 — costly to extend or operate for non-JVM teams
- The README itself notes a smaller community than OpenClaw or Hermes Agent, with a thinner ecosystem
- The runtime does not promise exactly-once behavior for external side effects; payments, sends, and publishes still need provider idempotency or review
- The optional DSH runtime requires installing and managing the dsh-rpc-agent child process — a separate operational surface
- Database options are fragmented (H2 dev, PostgreSQL 16 Docker default, MySQL supported, Kingbase opt-in), so multi-environment migration requires validating Flyway scripts yourself
How do you install or deploy this agent?
From source: git clone https://github.com/mateaix/mateclaw.git; backend: cd mateclaw-server && mvn spring-boot:run (http://localhost:18088); frontend: cd mateclaw-ui && npm install && npm run dev (http://localhost:5173); default login admin/admin123. Docker: cp .env.example .env && docker compose up -d (http://localhost:18080). Desktop: download the Electron binary from GitHub Releases — bundles JRE 21, no Java install needed. Requires Java 21+.
How do you use this agent?
Log in with admin/admin123; in Settings → Models drag providers into priority order and enter API keys; create digital employees (six built-in templates: General Assistant, Product Assistant, Research Analyst, Customer Support, Data Analyst, Code Reviewer) and bind skills, MCP tools, and knowledge bases; monitor every employee's runtime, tokens, and force-recycle stuck runs in Settings → System → Runtime; configure DingTalk/Feishu/Telegram channels, where magic commands /new, /clear, /status, /stop, /model, /help work; use Personal Access Tokens for headless scripts and CI.
How does this agent compare with similar options?
The README compares MateClaw with OpenClaw (TypeScript, config-file first, single-user), Hermes Agent (Python, single-user CLI), Claude Code (proprietary, Anthropic-only) and Cursor (IDE-only): MateClaw's differentiators are team-oriented multi-user RBAC/approvals/audit, LLM Wiki citation-backed knowledge digestion, and the Java/Spring Boot stack with five delivery surfaces; OpenClaw and Hermes Agent fit one-user-one-laptop scenarios better.