Hi~大家好，我是三金。
我最近在使用 VibeKankan 的时候发现它的版本已经很久不更新了！打开它的 Github 一看，竟然要停止维护了，最新版本还是我现在使用的 0.1.44，已经半年没更新过版本了！
最近一次的 commit 还是3 周以前，这就很难受，好担心它后面都没人维护。也想过直接拉源码改，但最终还是想拿来即用。故而又找到了另外一个平台——Multica。
能做什么
简单来说，这是一个可自定义多 Agent 协作的平台。
官方的定位 Multica 是一个人类和 AI Agent 共同工作的团队工作区，我们可以：
- 自定义各种 Agent；
- 人做中枢或者指定 AI Agent 做中枢，给其他 AI 同事派活；
- 同一个团队里，上下文共享，工作里的讨论、记录以及产出都会留在工作区；
- 如果已经有成熟的 AI Native，我们甚至可以将多个 Agent 组成小队，然后实现任务下发后的自动化开发；
- 根据会话和 memory 自动沉淀团队经验，将可重复的做法收敛成 skill，然后给多个智能体复用。
以下是「工作」菜单下的任务、项目以及自动化的截图：
使用
使用官方 Online 的话有两种模式：Web（multica.ai）以及桌面客户端 Multica Desktop，支持邮箱注册登录或者 Google 一键登录。
客户端的优势是登录后自动连接当前电脑，而 Web 端则需要通过命令行进行连接。
三金没有选择官方 Online 的方式来玩儿，而是通过自部署 + 客户端的方式启动。
自部署以及托管
自托管的 Multica 服务，主要运行 Web、API 和 PostgreSQL，我们可以通过 Docker Compose 快速启动。
另外，电脑上还需要：Git、Mack、curl 和 OpenSSL。这些前置条件准备好之后，我们就可以进行自部署了。
git clone --depth 1 https://github.com/multica-ai/multica.git
cd multica
make selfhost
执行 make selfhost 的作用 会从 .env.example 创建 .env，然后随机生成 JWT_SECRET、PostgreSQL 密码和 MULTICA_VCS_SECRET_KEY（自托管 Git 集成的加密密钥）。 另外，PostgreSQL 以及 Multica 的前端和后端镜像也会随之拉取。
上述任务结束之后我们可以通过以下两个操作来确认服务是否已就绪：
# 查看容器状态
docker compose -f docker-compose.selfhost.yml ps
# 检查 backend、数据库以及 migration
curl -fsS http://localhost:8080/readyz
# 正常响应是 {"status":"ok","checks":{"db":"ok","migrations":"ok"}}
确认 ready 之后，我们就可以通过 http://localhost:3000 来访问了：
三金这里主要使用桌面应用，从 Multica 下载页 下载对应系统的安装包进行安装，然后使用和 Web 版相同的邮箱进行登录。登录后，Desktop 会启动自己内置的 Multica CLI，并检测本机已有的 AI 编程工具。
Multica 默认是没有配置邮件服务的，所以我们输入自己的邮箱以后，需要从 backend 容器日志中进行读取：docker compose -f docker-compose.selfhost.yml logs backend | grep "Verification code"，然后在日志中会出现类似这样的内容：[DEV] Verification code for you@example.com: 123456
那如何配置邮件服务呢？可以通过在配置文件中设置主机和发件地址（以下只是引导，需要大家填写真实的配置信息），然后重启容器来激活邮件功能：
SMTP_HOST=smtp.example.com
SMTP_PORT=587
SMTP_USERNAME=multica
SMTP_PASSWORD=<password>
SMTP_FROM_EMAIL=noreply@example.com
快速上手
登录之后的第一件事，就是要创建工作区：
以 AI-TEAM 为例，创建好之后视图如下：
我们先点击运行时来看下当前机器上都有哪些工具：
然后回到智能体中，创建新的智能体：
接下来让我们来测试一下，输入提示词：在 Blender 中制作一只豚鼠。
这里用的 glm-5.3-flash 模型，形状是有的，但是有一种石质感，可能换成 gpt-6.1-sol 模型就好了。
有一说一，3D 建模的时间很长，足足一个多小时将近两个小时。
创建多 Agent
在测试单 Agent 成功以后，我们就可以尝试创建一个 Agent 团队出来了，比如搞一个专门做 3D 小游戏的团队：Game Designer + 3D 建模专家 + Game Development + Game QA & Playtest。
在小队菜单里将这几个 Agent 添加进去，3D 开发团队就这样诞生了（但是能不能投入使用肯定是需要很长的摸索迭代）：
组团打怪
我们新建一个任务，然后选择对应的团队：
接下来就是等待时间，先由 Game Designer 进行设计，然后再让 3D 建模专家进行 3D 建模，再将建模素材转交给开发 Agent 进行开发（不过后面没有执行 QA🤷♀️）：
经历了 1 个小时 16 分钟以后，一个划船游戏就完成了：
看板
任务菜单这里，就是一个进度看板，在这里会展示「待规划\待办\进行中\审核中\已阻塞\已完成」6 个任务状态。
可以看到，我们在上一步中完成的长船之旅的任务目前已经到了审核阶段，用鼠标进行任务卡片拖动或者在任务详情中更换状态，都可以更新该任务到下一个阶段：
定时任务
这是目前市面上大多数 AI Agent 客户端都有的能力：
包括 WorkBuddy、ZCode、Codex 这些，这里就不过多赘述了。
有需要做多 Agent 以及 AI 任务面板能力的小伙伴，可以试试 Multica。
Github：github.com/multica-ai/… star 哦～
官网：multica.ai
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Hi. Hi. Hi.
When I recently used VibeKankan, I found that its version had not been updated for a long time! Turn it on, Github, and look at it. It's still the latest version of 0.1.44 I'm using. It's been six months!
The last time I came was three weeks ago, it was hard, and I was so worried that nobody was keeping it. I was thinking about pulling the code directly, but eventually I wanted to use it. So I found another platform, Multica.
What can I do?
In short, it is a platform for customizing more Agent collaboration.
The official location of Multica is a team work area where humans and AI Agent work together, and we can:
- Customize various Agents;
- People act as hubs or designate AI Agent as hubs for other AI colleagues;
- In the same team, context is shared and discussions, records and outputs in the work area are left in the work area;
- If there is a mature AI Native, we can even create teams of Agents and then automate the development of the mission;
- Repetition of repeatable practices into skill and reuse of multiple AI agents, based on session and memory self-settling team experience.
Below is a screenshot of the tasks, projects and automation under the Work menu:
Use
Using official Online, there are two modes: Web (multica.ai) and desktop client Multica Desktop, which support mail registration or Google.
The advantage of the client is to automatically connect the current computer after login, while the Web end needs to connect via command line.
Sankin did not choose the official Online way to play, but started by deploying + client.
Self-deployment and hosting
Self-hosted Multica service, mainly running Web, API and PostgreSQL, we can start quickly through Docker Company.
In addition, there are computer requirements for Git, Mack, Curl and OpenSSL. Once these preconditions are ready, we can deploy ourselves.
git clone --depth 1 https://github.com/multica-ai/multica.git
cd multica
make selfhost
Executing make selfhost creates.env from.env.example and then randomly generates JWT SECRET, PostgreSQL passwords and MULTICA VCS SECRET KEY (auto-host Git integrated encryption key). Also, front-end and back-end mirrors for PostgreSQL and Multica will be pulled.
After the completion of the above-mentioned tasks, we can confirm service readiness through two operations:
# View container status
docker compose -f docker-compose.selfhost.yml ps
# Check backend, database and migration
curl -fsS http://localhost:8080/readyz
# The normal response is #
After confirming that we are ready, we can visit the following:
This is mainly a desktop application where you download the installation package of the corresponding system from the Multica download page and then log in with the same mailbox as the Web version. After login, Desktop will start its own embedded Multica CLI and test the AI programming tool available on this machine.
Multica defaults that no mail service is configured, so when we enter our own mailbox, we need to read it from the backend container log: docker log -f docker-compose. Selfhost.yml logs backend | grep "Verification code" and then similar content will appear in the log: [DEV] Verification code for you@example.com: 123456
How do we fit the mail service? You can activate the mail function by setting up the host and the sending address in the configuration file (the following is just a guide, you need to fill in the real configuration information) and restart the container:
SMTP_HOST=smtp.example.com
SMTP_PORT=587
SMTP_USERNAME=multica
SMTP_PASSWORD=<password>
SMTP_FROM_EMAIL=noreply@example.com
Let's do it fast.
The first thing after the login is to create a workspace:
To take the example of AI-TEAM, create a later view as follows:
Let's click on the running to see what tools are on the current machine:
Then back to AI agent, creating a new AI agent:
So let's take a test and enter prompt: Make a guinea pig in Blender.
The glm-5.3-flash model used here has a shape, but there is a sense of stone that could be replaced by a gpt-61-sol model.
One says that 3D modelling takes a long time, over an hour, nearly two hours.
Create MultiAgent
After the test sheet Agent is successful, we can try to create an Agent team, like a team dedicated to 3D games: Game Designer + 3D Modeller + Game Development + Game QA & Playtest.
Add these Agents to the team menu, so 3D development teams are born.
♪ The troupe ♪
We're going to create a new task, and we're going to select the team:
This is followed by a waiting period for the design of the Game Designer, followed by 3D modelling by 3D modelling experts, and the transfer of model asset to the development of Agent (but not later implemented QA🤷♀️):
After 1 hour, 16 minutes, a rowing game was completed:
Watch.
The task menu here is a progress board where you will display the six tasks "to be planned \\to-do\valued\completed".
As you can see, the task of a longboat trip that we have completed in the previous step is now at the audit stage, with a mouse tow the task card or change the status in the details of the task, which can be updated to the next stage:
Time job
This is the ability of most of the AI Agent clients on the market today:
That includes WorkBuddy, ZCode, Codex, which is not much to dwell on.
There are small partners who need to do more Agent and the AI task panel capability to try Multica.
Github: Github.com/multica-ai/...star
Official: multica.ai
Auto-fetched and translated by AI Atlas. All rights belong to the original author.