Introduction

When using a single agent in DSH, the common problem isn’t just “getting the model to answer,” but putting multiple roles, customer context, and approval chains into a single workflow. ai-company-framework is a DeepSeek Harness dsh.bundle designed to create a company composed of multi-role Agents with a single sentence, while allowing users to retain approval rights before major decisions.

Below is an introduction to its positioning, core capabilities, installation and enablement methods, and precautions.

What is this

ai-company-framework is a dsh.bundle that transforms a one-sentence requirement into a company composed of multi-role Agents.

Verified information includes:

  • Supports roles such as writer, editor, developer, QA, sales, customer service, etc.;
  • Major decisions are pushed to mobile phones via the Lark human decision gate; they do not land until the user approves;
  • The installation package does not contain user data; the database, logs, and credentials are created locally upon first use;
  • No RAG, no vector database, no telemetry.

Its declared version is 0.3.10 with an MIT license. Documentation confirms it has been verified on DSH Desktop @deepseek-ai/dsh 0.1.0-rc.8 (Windows).

Core Capabilities

Company Mode Disabled by Default

Company mode is disabled by default.

To enable:

/company

To disable:

/no-company

This means normal sessions are not affected; relevant capabilities are only triggered after entering company mode.

Multi-role Agent Company

It can create a multi-role Agent company with a single sentence; roles include writer, editor, developer, QA, sales, customer service, etc.

Built-in components include:

  • 15 skills gated by /company;
  • 7 workflow templates;
  • AgentTeams runtime;
  • web activity panel;
  • employee sidebar;
  • Lark bridge.

Customer Profiles & Auto-Merge

It establishes SQLite profiles per customer and automatically merges private chat and multi-group data for the same customer.

Customer profiles are isolated by ID.

Agent Push

Agents proactively push results upon completion, without relying on polling.

Lark Bot and Human Decision Gate

Major decisions are pushed to mobile phones via the Lark human decision gate; they do not land until the user approves.

It can also register a Lark bot for the boss and any employee via a single scan code (registerApp).

Isolation & Security

Verified isolation and security points include:

  • Sidebar state is isolated by session;
  • Model reconfiguration has 403 ownership checks;
  • Customer profiles are isolated by ID;
  • Lark App Secrets use Windows DPAPI (CurrentUser scope) and are saved locally only;
  • Installation package has zero user data;
  • Database, logs, and credentials are created locally upon first use;
  • No postinstall hooks;
  • Installation and uninstallation are managed by dsh plugin.

Installation & Enablement

First, prepare the ai-company-framework-0.3.10.tgz package, then run in the terminal:

dsh plugin --profile web add .\ai-company-framework-0.3.10.tgz

After installation, enable company mode:

  1. Restart DSH Desktop.
  2. Open a new session.
  3. Run:
/company

If you want to exit company mode later, run:

/no-company

Typical Usage

After enabling /company, you can create a company and the first order with a single sentence. The example in the documentation is:

I want an ecommerce content company — first order: a Xiaohongshu post + customer-service scripts.

Subsequent major decisions will be pushed to the mobile phone via Lark and will only be executed after approval.

Suitable Scenarios & Notes

It is suitable for users who need to organize multi-role Agent collaboration in DSH, consolidate context by customer, and perform approval and bot interaction via Lark.

Since it explicitly has no RAG, no vector database, and no telemetry, it is also suitable for scenarios focusing on local data boundaries and dependency scope.

The plugin runs with the permissions of the current DSH process; you should check the source code and license before installing. The license is MIT, and the version is explicitly 0.3.10 in the documentation. Documentation states it has been verified on DSH Desktop @deepseek-ai/dsh 0.1.0-rc.8 (Windows).

Links

  • Community Directory Page: https://www.skillhub.cn/plugins/yyyy231209/ai-company-framework
  • GitHub Repository: https://github.com/yyyy231209/ai-company-framework