MateClaw builds AI coworkers on Spring AI Alibaba
The open-source project packs agent teams, MCP tools, skills and long-term memory into one Java runtime. Version 2.2.0 requires Java 21.
In short
MateClaw is an open-source Java runtime that runs several AI agents as a team, gives them MCP tools, skills and persistent memory, and reaches users through eight messaging channels.
At a glance
- Version 2.2.0 is dated August 29, 2026; Apache 2.0 license, about 1,100 stars and 332 forks on GitHub.
- Built on Spring Boot 3.5 and Spring AI Alibaba 1.1, Java 21 or newer, with a Vue 3 front end.
- Nine model providers wired in: DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio and MLX.
- Eight IM channels: DingTalk, Feishu, WeChat Work, WeChat, Telegram, Discord, QQ and Slack.
- Every figure comes from the project's own description of itself; no independent verification is available.
MateClaw is an open-source Java runtime for AI agents that work a job as a team rather than one prompt at a time. It sits on Spring Boot 3.5 and Spring AI Alibaba 1.1 and ships under Apache 2.0. The current release, 2.2.0, is dated August 29, 2026 in the project's release list. The repository shows roughly 1,100 stars and 332 forks.
Agents with a personnel file
The project frames its agents as digital coworkers with a role, a goal and a backstory instead of a blank chat box. Six templates ship with it, ranging from research analyst to code reviewer. Identity and state live in plain files in the workspace — SOUL.md, PROFILE.md, AGENTS.md and MEMORY.md. The stated aim is that an agent survives a backend restart with its history intact.
One runId per job
A Team Run ties the objective, the task graph, each worker's execution and the final deliverables to a single durable runId. A shared board with eight statuses hands work out in parallel and resolves prerequisites through a blockedBy field. Delegation to subagents is capped at three levels deep. Chat, a live view and the team history all read the same underlying record.
Tools, skills and a wiki
Tools arrive through the Model Context Protocol over stdio, SSE or HTTP and are bound per agent. Skills are packages in the SKILL.md format carrying a manifest, a prompt, a tool list and a LESSONS.md file. An 'LLM Wiki' turns source material such as PDFs and scraped pages into linked entries that keep their citations. Nine model providers are wired in, among them DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio and MLX.
Where it runs
Users reach the agents through a web console, a desktop build, an embeddable widget and eight messengers: DingTalk, Feishu, WeChat Work, WeChat, Telegram, Discord, QQ and Slack. For storage the README lists H2 for development plus PostgreSQL 16, MySQL 8.0 or newer and Kingbase. A local server comes up on port 18088 and the Docker stack on port 18080. Java 21 or newer is required.
What we could not check
'Dream' sits in the repository's own tagline, yet the README only sketches it as post-conversation extraction plus scheduled consolidation, without spelling out the steps. Everything above comes from the project's own account of itself. No independent second source on MateClaw was available at the time of writing. That leaves the central question — how well the orchestration and the memory hold up in sustained use — unanswered.
FAQ
What is MateClaw?
An open-source Java runtime in which several AI agents work a task as a team, call tools over MCP, and keep their memory in plain files inside the workspace.
Which AI models does MateClaw support?
The project lists nine providers: DashScope, OpenAI, Anthropic, Gemini, DeepSeek, Kimi, Ollama, LM Studio and MLX. Local models via Ollama, LM Studio and MLX are explicitly covered.
What do you need to self-host MateClaw?
Java 21 or newer plus a database — H2 for development, or PostgreSQL 16, MySQL 8.0 or newer, or Kingbase in production. The server runs locally on port 18088 and the Docker stack on port 18080.
