Building a Modern AI Workbench: Codex + Feishu CLI with 27 Office Skills

Introduction

The author, a long‑time Feishu user, attended the Feishu Demo Day and noticed that several speakers were using approaches similar to his own: combining Codex with the Feishu CLI to automate office tasks.

He also references other speakers who demonstrated AI‑driven email and document handling, and mentions his own open‑source projects z-mail-reader, z-smart-xparse and z-md-excel that target these scenarios.

Codex Plugin for Feishu

The author created a Codex plugin that wraps the Feishu CLI and exposes the 27 official Feishu Skills. The plugin’s entry point is the command larksuite-cli. The most frequently used Skills are listed below: lark-doc – document handling lark-base – multi‑dimensional tables lark-sheets – spreadsheets lark-whiteboard – whiteboard lark-mail – email lark-vc – video conference lark-im – instant messaging

After installing the plugin, the first step is to initialise a Feishu app as prompted by the CLI (see image).

Typical Workflows

1. Codex Automation + Feishu Instant Messaging

The author previously used Codex to schedule periodic checks for OpenAI and Claude updates, sending notifications via Bark to his phone. By switching to the Feishu plugin, the lark-im Skill can send and reply to messages, search chat history, manage group members, and upload/download files (including large‑file chunked download).

This enables AI‑generated summaries to appear instantly in Feishu groups on the author’s phone, eliminating the need to open a browser.

2. Email Processing Pipeline

The author subscribes to many newsletters and receives daily work emails, often with PDF attachments. He uses the three open‑source tools mentioned earlier to automate the pipeline: z-mail-reader fetches emails and attachments. z-smart-xparse parses PDFs, images, and scans into structured Markdown. z-md-excel converts Markdown tables into Excel files.

The extracted data is then written into Feishu Base for review, collaboration, statistics, and tracking. A diagram (see image) illustrates the flow from email receipt to structured data in Base.

For large documents, such as a 300‑page tender, the author runs z-smart-xparse to split and parse the PDF, then uses an AI Agent to extract qualification requirements, scoring rules, and delivery risks. The full analysis is stored in Feishu Docs, while a risk list resides in Feishu Base, where each item can be assigned an owner, status, and deadline.

3. Document Generation via Whiteboard

A command chain demonstrates how a PDF can be turned into an editable PPT infographic:

lark-whiteboard generate board → lark-doc write document → embed HTML in Feishu Doc → send to Feishu group

The author shows an example where a PDF is fed to Codex, which creates a live PPT infographic on a Feishu whiteboard; the result can be edited directly.

Future Workbench Vision

The author argues that Feishu is an ideal AI workflow operating system: Codex, Claude Code, and various Agents perform reasoning and execution, while Feishu stores the results in collaborative spaces.

Key capabilities he highlights are:

Reading real‑world materials

Invoking real tools

Persisting results in places the team accesses daily

Supporting replay, review, traceability, and hand‑off

These workflow features, rather than raw model capability, differentiate truly useful enterprise AI. This is why he built the Feishu CLI as a Codex plugin.