Multica

开源的多智能体团队协作平台，将 AI 编程助手转变为团队成员

Vibe Coding

https://github.com/multica-ai/multica

Visit Website

In-depth Report

Multica is an open source multi-agent team collaboration platform that is dedicated to integrating mainstream AI coding agents such as Claude Code, Codex, OpenCode, OpenClaw, and Gemini into a unified task management and execution system, allowing AI Agents to work like real team members. The project was open sourced on GitHub on January 13, 2026. As of April 17, 2026, it has accumulated more than 15,400 stars and 1,900 Forks, becoming one of the fastest growing open source AI Agent management platforms on GitHub in just three months.
Developed by an independent team, Multica is a vendor-neutral open source project that supports access to a variety of mainstream AI coding tools. The project adopts a self-hosted design, and enterprises can deploy services on their own infrastructure to ensure data privacy and security. In the field of AI Agent, although a single AI programming assistant is powerful, it is always a "go it alone" mode. The emergence of Multica fills a gap in the market. It provides a unified management and collaboration platform for multiple AI assistants.
The first core function is task allocation and progress tracking. Users can create Issues on Multica and assign them to AI employees just like assigning tasks to colleagues. AI will automatically receive tasks, execute code, report progress, comment under tasks, and create subtasks. The second core feature is multi-AI access support. Multica supports access to a variety of mainstream AI coding tools, including Claude Code, Codex, OpenCode, OpenClaw, Hermes, etc. The vendor-neutral design ensures flexibility. The third core function is skill accumulation. When AI solves a complex problem, its solution methods and ideas can be precipitated into reusable skills for use by other AIs or subsequent tasks. The fourth core function is real-time progress push. Progress is pushed in real time through WebSocket, and AI reports work status like colleagues. In terms of technical architecture, Multica adopts the technology stack of Next.js + Go + PostgreSQL.
Multica is an open source free project with a custom open source license. Users can deploy and use it on GitHub for free. Business models may include: enterprise version providing more powerful management capabilities and professional technical support; hosting services providing SaaS deployment options; custom development services to meet specific needs.
Judging from community feedback, Multica has received positive reviews. The main advantages include: innovatively treating AI as a team member rather than a tool; the Task-Board interface is intuitive and easy to use; multi-AI access is flexible; and the skill accumulation mechanism is valuable. Potential problems include: as a new project, the documentation and ecology are still being improved; technical background is required for deployment and maintenance; and there are certain requirements for team collaboration processes.
The emergence of Multica fills the gap in AI Agent management platforms. Compared with deep research frameworks such as DeerFlow, Multica focuses more on multi-agent collaboration in programming assistance scenarios. The two are complementary rather than directly competitive. The technical community has paid great attention to Multica, and Zhihu, Nuggets, CSDN, Tencent Cloud Developer Community and other platforms have in-depth analysis articles.
New project risks: As a new project in early 2026, long-term maintainability remains to be seen. Learning curve: Need to understand and adapt to new models of human-machine collaboration. Technical threshold: Self-hosted deployment requires certain technical capabilities.
Multica is suitable for the following scenarios: software development teams who want to introduce AI members; companies that need to manage multiple AI programming assistants; developers who have organized needs for AI development workflows. Not recommended scenarios: It is enough for individual developers to use a single AI assistant; there is no support from a technical team to maintain the system.
Multica is an innovative human-machine collaboration platform that transforms AI from tool to team member. Although it is a new project, its core concept and growth rate are worthy of attention. As AI agents play an increasingly important role in development work, collaboration platforms like Multica may have broader room for development.

User Reviews

Christo_pherRoss

—

试了一下,确实比开五个终端窗口强多了。Agent跑完了会自己更新状态,不用我去轮询。

KylePerez520

—

之前用Claude Code写代码,每次都要在终端里盯着它跑,中间还不能切去做别的。现在把任务丢给Multica,它在后台跑,我该干嘛干嘛,跑完了会通知我。这种从「保姆」到「管理者」的角色转变,用过的都懂。

黄雪浩

—

用了两周,回不去了。

Donald_Phillips0071

—

豆瓣没评论,我来这里说一句:Daemon不能断,断了Agent就失联,建议tmux伺候。

KimberlyRuiz_20208

—

用了三周,实际体验下来有几个点想补充。第一,技能复用确实是个好功能,但还没有宣传的那么成熟——目前更像是个高级版的issue template,还不是真正的「能力复利」。第二,多Agent并行的时候偶尔会出现调度冲突,两个Agent被分配到同一个代码库的不同任务时,代码冲突还是要人来解决。第三,自托管部署对技术能力有要求,小白建议直接用Cloud版。总的来说,方向正确,但还处于PMF早期。

薛明阳

—

深度用了三周,说点真实感受。先说好的:第一,Agent as teammate这个理念执行得很到位——看板、assignee、comment、status,Agent在这些协作界面上的出现方式跟人类成员完全一致,上手非常自然。第二,Runtime抽象做得好,一台MacBook Pro、一台Linux服务器、一台ARM云实例,都能注册成Runtime,统一调度。第三,Squad机制让多Agent编排有了雏形,虽然不是完美的编排引擎,但在v0.x版本已经够用了。说说不足:License是改过的Apache 2.0,不是纯开源,商业化前要仔细看条款。Skills系统目前还是基于关键词匹配,不是真正的向量语义检索,pgvector基本处在「已部署未启用」的状态。另外文档虽然写得不错,但部分操作细节(比如自托管版本更新流程)还是靠猜。总的来说,作为项目管理型的Agent平台,Multica是目前最成熟的选项,但距离「AI-native production system」还有距离。

Samantha_Green520

—

我是做后端开发的,团队里三个人加五个Agent并行开发。以前最头疼的是上下文管理——Agent A修完的bug,Agent B不知道,又踩一遍。现在Multica把技能沉淀下来,A学过一次的东西B也能用,确实省了不少重复劳动。

MaryGomez_88

—

看板加Agent,绝了。

RFosterJr

—

产品定位很清晰,就是给已经用Claude Code这些工具的人用的,小白慎入。

袁云

—

我是独立开发者,一个人管五个Agent,Multica对我来说简直是神器。以前我得在五六个终端窗口之间来回切,记不住哪个Agent在干什么。现在一块看板全看完,谁在跑、谁空闲、谁卡住了,一目了然。而且Agent跑完了自己更新状态,我不用一直盯着。省下来的精力可以专注在更需要创造力的部分,比如架构设计。

LoganMartin

—

Squad的Leader Agent分配机制挺有意思。之前我得手动判断每个任务分给谁,现在直接丢给Team,它自己调度。虽然偶尔分配不太合理,但大部分时候比我自己分的还准。

David_Jenkins007

—

安装还挺简单的,brew一行搞定。

DRcla

—

Squad功能不错,把小队长设好之后,任务直接@前端组就行了。

purplemouse387

—

试了几个小时,我个人的感觉是:如果你只是偶尔用AI写个脚本,没必要上Multica。但如果你同时管着好几个Agent,每天开五六个窗口来回切,那它确实能救命。门槛是有,但值得花那个时间。

LubbertNeuteboom

—

对比了一下同类产品:Vibe Kanban更适合单人在本地跑多个Agent,Paperclip想做的是把人类移出流程的无人公司,而Multica走的是折中路线——人和Agent在同一块看板上协作。三种思路各有优劣。我选Multica的原因很简单:我既不想当保姆(人的精力有限),也不放心完全交给AI自己跑(现在的Agent还没那么可靠)。中间路线最适合当前的阶段。

NancyJackson62

—

作为一个技术负责人,我来聊聊Multica在团队里的落地情况。我们团队6个人,之前用Claude Code完全是散养状态——每个人自己开终端窗口,谁也不知道别人的Agent在干什么。引入Multica之后,最大的变化不是效率提升(当然也有),而是可见性。现在我在看板上能看到每个Agent的状态、进度、阻塞点,甚至能追溯一周前的执行记录。这对于我们这种强调「团队协作」而不是「个人英雄主义」的团队非常重要。但也遇到了几个问题:第一,不是所有人都愿意改变工作习惯——有人觉得直接敲命令行更快。第二,团队规模大了之后,权限管理还不太够用。第三,Agent执行结果的质量参差不齐,还是需要人工review。我的建议是:先在一个项目组试点,跑通了再推广,不要一上来就全面铺开。

EliKristensen

—

终于不用再盯着终端了。

RaymondRivera_998

—

自托管党狂喜,Docker一把梭,数据不用出自己服务器。

JJohnson_77

—

看了Jiayuan的访谈,他之前的Devv和DevCode都失败了,第三次才找到PMF。挺佩服这种打法的——不是去造更聪明的Agent,而是做Agent的管理层。这个切入角度确实刁钻。

AlexanderHarris_9965

—

技能复用这个设计很妙。

Rachel_Murphy_20226

—

看完硅星人那篇深度报道才发现Multica背后的故事挺有意思。创始人Jiayuan之前做的Devv(开发者搜索引擎)和DevCode(AI编程)都失败了,第三次创业才找到这个方向——不造Agent,而是做Agent的管理层。他说了一句话很触动我:产品功能和技术本身已经没壁垒了,Coding已经被商品化了。确实,现在拼的不再是模型能力,而是怎么把AI真正嵌入到工作流里。

ul1xsvrnzw

—

那个andrej-karpathy-skills引流操作是真的骚,但确实有效。

AFoster

—

我觉得最大的短板是编排能力还太弱。目前基本上是人来拆任务,Agent只管执行。如果真的要做到复杂工作流自动编排,还有很长的路。不过作为v0.x版本,方向是对的。

Gregory_Mendoza_2024

—

跟GitHub Issue #815那个深度分析的观点基本一致。Multica目前最大的问题不是功能不够多,而是它本质上还是在「用管理人的方式管理AI」。Issue的状态流转是扁平的——backlog、todo、in_progress、in_review、done、blocked、cancelled,这些状态跟人用的Jira没区别。但AI的工作流跟人的工作流本质上是不同的——AI可以同时处理多个任务、可以24小时不间断运行、不需要休息。把AI塞进为人类设计的状态机里,其实是一种「削足适履」。不过我理解这是团队打的「安全牌」——先做人类熟悉的东西,再逐步演进到AI原生。作为MVP,Mutica的方向是对的,但我很期待看到它什么时候能真正从「AI teammate manager」进化到「AI orchestrator」。

WolfWalletChavez

—

感觉像雇了一群免费牛马。

EugeneHicks_Max5

—

装完发现Codex支持不太好,官方说支持实际有坑,后来切Claude Code就好多了。

Garrett294

—

说几个问题吧。第一,License不是纯Apache 2.0,是改过的,加了商业限制,采购前要看清楚。第二,标称支持8种CLI,实际第一梯队只有Claude Code和Codex,其他的稳定性参差不齐。第三,Skills系统还没有向量记忆,检索靠人手工挂载。第四,624个open issues说明还在快速迭代期,生产环境使用需谨慎。但作为开源项目,迭代速度和质量在同体量项目里算很不错了。

月光_11

—

用Docker自托管部署的,过程还算顺利,就是版本更新太频繁了,几乎每天都有新release。好处是迭代快bug修得快,坏处是我隔两天就得pull一次新镜像。

Betty.GomezII25

—

623个open issues,说明还在快速迭代。

zOEYoLSEN

—

部署的坑:如果你选的海外服务器,Agent调用大模型API延迟会低很多。我一开始用的国内服务器,Claude Code经常超时。后来切到香港节点就好多了。另外API Key要提前在对应的CLI里配好,不是配在Multica里,这个文档没说清楚,很多人在这上面卡住。

Multica

In-depth Report

User Reviews

Similar Products