lark-cli-mcp
# @yoreland/lark-cli-mcp
> One-line `npx` MCP server that lets AI clients (Amazon Quick Desktop, Claude Desktop, …) operate **Feishu / Lark as your own user identity** — send, read, reply, and search messages.
It wraps the official [`lark-cli`](https://github.com/larksuite/cli) (bundled as a dependency, no separate install) and exposes **19 tools** over MCP stdio — messaging, cloud docs, Wiki, Drive, and Bitable. Because every call runs with `--as user`, the sender/creator shown in Feishu is **you**, not a bot.
---
## Quick start (for workshop attendees)
### 1) One-time login (in a terminal)
```bash
# Bind the shared Feishu app (scan the QR code)
npx -y @yoreland/lark-cli-mcp -- config init --new
# OAuth device-flow login as yourself
npx -y @yoreland/lark-cli-mcp auth
# Sanity check (should report logged-in)
npx -y @yoreland/lark-cli-mcp doctor
```
### 2) Add the MCP server in Amazon Quick Desktop
Go to **Settings → Capabilities → MCP → + Add MCP**:

In the dialog, use **Paste JSON config** and paste:
```json
{
"command": "npx",
"args": ["-y", "@yoreland/lark-cli-mcp"]
}
```

Click **Apply** → **Save**. The server should show **19 tools · Connected** ✅
> Manual form instead of JSON? Connection type = `Local`, Command = `npx`, Arguments = `-y @yoreland/lark-cli-mcp`. Arguments are space-split (no spaces inside the package name, so this is safe).
### 3) (Recommended) Add the skill
**Settings → Capabilities → Skills → Upload**, then select [`skill/feishu-lark/SKILL.md`](skill/feishu-lark/SKILL.md). Toggle it **Active**.

The skill teaches Quick *when* and *how* to use the 19 tools (resolve chat_id/open_id first, confirm before sending, Bitable field handling, etc.) — the MCP provides the tools, the skill makes Quick use them well.
### 4) Try it
- "看看 XX 群最近聊了什么"
- "帮我找一下某人"
- "搜一下我的飞书文档里有没有 …"
- "列一下这个多维表格 \<URL\> 的记录"
---
## Prerequisites
- **Node.js v18+** (`node -v`)
- An MCP client (Amazon Quick Desktop / Claude Desktop / etc.)
- A Feishu/Lark account in the org running the workshop
- Network access to npm registry and Feishu OAuth
The Feishu **App ID / Secret** is provided centrally by the workshop host and baked into the shared `lark-cli` config — attendees only do the OAuth login step. (See [Host setup](#host-setup).)
> Behind the Great Firewall? Use a mirror: `npm config set registry https://registry.npmmirror.com`
---
## Commands
```bash
npx -y @yoreland/lark-cli-mcp # start MCP server (stdio) — what the client runs
npx -y @yoreland/lark-cli-mcp auth # OAuth device-flow login (user identity)
npx -y @yoreland/lark-cli-mcp status # show auth status
npx -y @yoreland/lark-cli-mcp logout # clear token
npx -y @yoreland/lark-cli-mcp doctor # environment + login self-check
npx -y @yoreland/lark-cli-mcp -- <args> # passthrough to bundled lark-cli
```
---
## The tools
### Messaging (IM)
| Tool | What it does |
| ------------------------ | ----------------------- |
| `feishu_send_message` | Send a message |
| `feishu_get_messages` | Read recent messages |
| `feishu_reply_message` | Reply (thread optional) |
| `feishu_search_messages` | Search messages |
| `feishu_list_chats` | Find group chats |
| `feishu_search_user` | Find a user (→ open_id) |
| `feishu_get_thread` | View a thread |
### Docs / Wiki / Drive
| Tool | What it does |
| ----------------------- | ---------------------------------- |
| `feishu_search_docs` | Search docs / wiki / sheets |
| `feishu_doc_fetch` | Read a document |
| `feishu_doc_create` | Create a document (markdown) |
| `feishu_doc_update` | Update a document |
| `feishu_drive_search` | Search Drive files (type filters) |
| `feishu_wiki_node_list` | List wiki nodes |
| `feishu_wiki_node_get` | Get a wiki node (accepts URL) |
### Bitable (multi-dimensional tables)
| Tool | What it does |
| --------------------------- | ------------------------------------- |
| `feishu_base_table_list` | List tables in a base |
| `feishu_base_field_list` | List fields of a table |
| `feishu_base_record_list` | List records (filter/sort) |
| `feishu_base_record_search` | Search records |
| `feishu_base_record_upsert` | Create/update a record |
### Talk to it naturally
| Goal | Say to your assistant |
| --------------- | ------------------------------------------- |
| Read a group | "看看 XX 群最近聊了什么" |
| Send | "在 XX 群说:明天会议改到 3 点" |
| Reply | "回复那条消息:收到,我来跟进" |
| Search | "搜一下谁提过客户报价" |
| Find someone | "帮我找一下张三的 open_id" |
| View a thread | "看看那条消息下面的讨论" |
---
## Host setup
The workshop host creates **one** Feishu custom app and configures it so attendees share the same App ID/Secret but each authorize their own account.
1. [Feishu Open Platform](https://open.feishu.cn) → create an internal custom app → note **App ID / App Secret**.
2. Enable **User token scopes** matching the `im`, `contact`, `search` domains (message read/write, reply, chat read, user search, message search).
3. Distribute the App ID/Secret to attendees via `lark-cli config` (or a pre-bound config). The login step requests scopes via `--domain im,contact,docs,wiki,drive,base`.
`auth` uses **OAuth Device Flow**, so no `redirect URL` / `localhost:3000` callback configuration is required.
---
## Troubleshooting
**`missing required scope(s)`** — re-login with the needed domain:
```bash
npx -y @yoreland/lark-cli-mcp auth --domain im,contact,docs,wiki,drive,base
```
**Client shows "No tools loaded"** — run `npx -y @yoreland/lark-cli-mcp doctor`; confirm Node ≥18 and that `auth status` is OK.
**Token expired** — just re-run `auth`.
---
## Known limitations
- No image/file attachment sending (text + markdown only)
- No interactive cards
- No group creation
- Tokens expire; re-run `auth` when they do
---
## How it works
```
MCP client (Quick Desktop / Claude Desktop)
│ stdio (MCP / JSON-RPC)
▼
@yoreland/lark-cli-mcp (server.mjs)
│ child_process.execFile (no shell → injection-safe)
▼
lark-cli --as user (bundled dependency)
│ OAuth user_access_token (device flow)
▼
Feishu / Lark Open API
```
## License
MIT
TDQS
Scored across 7 tools
Each tool targets a distinct action: sending, replying, retrieving messages, listing chats, searching messages, searching users, and viewing threads. No two tools have overlapping purposes, making it easy for an agent to select the correct one.
All tools follow a consistent 'feishu_verb_noun' pattern (e.g., feishu_send_message, feishu_list_chats). This uniform convention aids predictability and reduces cognitive load.
With 7 tools, the server is well-scoped for a messaging-focused MCP server. Each tool addresses a core function without redundancy, and the count is right-sized for the domain.
The set covers essential messaging operations: send, reply, retrieve messages, search, and list chats. Minor gaps exist (e.g., creating chats, marking messages as read), but the core workflow is fully supported.