larksuite/cli

> As far as I know, Lark doesn't have an official MCP server for Claude Code yet？

据我所知，基本是第三方开发的 mcp，

公司禁止导出任何文档，审批直接自动拒绝。所以这里能不能自定义呢

任何lark-cli命令前加上http_proxy=xxxx

> > > One important advantage of a CLI is progressive context disclosure. While MCP, Skills, or typical HTTP calls often require sending a relatively complete context in each request, a CLI lets you incrementally reveal only what’s necessary at each step. This not only improves control over execution flow, but also significantly reduces token usage and overhead in LLM-driven workflows.
> >
> >
> > Is the incremental disclosure feature you mentioned the same as the incremental disclosure feature of Skills?
>
> Not exactly.
>
> With MCP/Skills, every time the LLM connects, it still needs to reason about whether to call a tool and how to call it. That decision-making process itself introduces extra context and token overhead.
>
> A CLI, on the other hand, doesn’t require that upfront reasoning loop. It exposes capabilities progressively — for example via --help or command-specific introspection — only when they’re actually needed.
>
> So you can achieve a similar outcome, but with m...

> > One important advantage of a CLI is progressive context disclosure. While MCP, Skills, or typical HTTP calls often require sending a relatively complete context in each request, a CLI lets you incrementally reveal only what’s necessary at each step. This not only improves control over execution flow, but also significantly reduces token usage and overhead in LLM-driven workflows.
>
> Is the incremental disclosure feature you mentioned the same as the incremental disclosure feature of Skills?

Not exactly.

With MCP/Skills, every time the LLM connects, it still needs to reason about whether to call a tool and how to call it. That decision-making process itself introduces extra context and token overhead.

A CLI, on the other hand, doesn’t require that upfront reasoning loop. It exposes capabilities progressively — for example via --help or command-specific introspection — only when they’re actually needed.

So you can achieve a similar outcome, but with much tighter control over co...

> One important advantage of a CLI is progressive context disclosure. While MCP, Skills, or typical HTTP calls often require sending a relatively complete context in each request, a CLI lets you incrementally reveal only what’s necessary at each step. This not only improves control over execution flow, but also significantly reduces token usage and overhead in LLM-driven workflows.

Is the incremental disclosure feature you mentioned the same as the incremental disclosure feature of Skills?

比较好奇开发者自己是怎么测试的，字节内部能接受 bot 权限这么大吗

from ai

> 搞定了。卸载重装后 lark-cli 已经可以正常运行了，版本 1.0.0。
>
> 之前的问题就是第一次没代理时安装失败，npm 留下了不完整的包（symlink 指向旧的
> run.js，但二进制文件没下全），导致权限异常。卸载后干净重装就没事了。

One important advantage of a CLI is progressive context disclosure.
While MCP, Skills, or typical HTTP calls often require sending a relatively complete context in each request, a CLI lets you incrementally reveal only what’s necessary at each step.
This not only improves control over execution flow, but also significantly reduces token usage and overhead in LLM-driven workflows.

+1 默认权限太多了且无法修改 希望能自己调整权限 或者使用已有机器人的权限

+1

> 按照配置步骤，一个授权 直接提了个审批到我leader那

这不合理，config init的时候，我已经把权限申请和审判好了，auth不应该修改，而应该是仅授权。

按照配置步骤，一个授权 直接提了个审批到我leader那

+1。在执行lark-cli config init进行配置的时候，需要认证。我使用私有化的飞书扫描登录的二维码，无法登录，显示“暂时无法访问”。

希望能尽快适配私有化飞书。

> > As far as I know, Lark doesn't have an official MCP server for Claude Code yet？
>
> https://open.larksuite.com/page/mcp [@debuggy](https://github.com/debuggy)

This link says: “Sorry, you don't have access to this page”, so maybe the mcp is not in GA stage yet?