DEV Community

Cover image for One Open Source Project a Day (No. 76): Knowledge Work Plugins - Anthropic's Official Role-Specialist Plugin Library
WonderLab
WonderLab

Posted on

One Open Source Project a Day (No. 76): Knowledge Work Plugins - Anthropic's Official Role-Specialist Plugin Library

#ai #opensource #mcp #claude

Introduction

"The best AI tool is the one that already knows your job."

This is the 76th article in the "One Open Source Project a Day" series. Today's project is Knowledge Work Plugins.

This one is different from most projects in this series — it's not from a third-party developer. It's from Anthropic itself. That alone is worth paying attention to: it tells us how Anthropic thinks about the concrete answer to "how does AI actually fit into knowledge work?"

One sentence: turn Claude into a specialist in your role, not just a general-purpose assistant. Eleven plugins covering sales, legal, finance, product management, marketing, data analysis, bio-research, and more — each with domain expertise, workflows, and direct connections to real tools. 15.5k Stars, Apache-2.0.

What You Will Learn

  • How the four-layer plugin architecture (Skills / Commands / Connectors / Sub-agents) is designed
  • Why fully declarative design (zero code, pure Markdown) is the most important architectural choice in this project
  • How MCP connectors let plugins talk directly to HubSpot, Snowflake, Figma, and other real tools
  • How to customize existing plugins for your company's processes, terminology, and toolchain
  • Why Anthropic chose to open-source these plugins, and what it signals about enterprise AI adoption

Prerequisites

  • Basic familiarity with Claude Code or Claude Cowork
  • Some experience in a knowledge work role (helps you appreciate each plugin's value)
  • Introductory understanding of MCP (Model Context Protocol) is optional but helpful

Project Background

Project Introduction

Knowledge Work Plugins is an officially maintained open-source library from Anthropic, designed for knowledge workers and compatible with both Claude Cowork (claude.com's collaborative workspace) and Claude Code (the developer CLI).

The project's core claim: a general-purpose AI assistant underperforms a domain specialist. The same question given to a Claude that has sales domain knowledge, is connected to a HubSpot CRM, and understands your company's product terminology, versus a bare Claude with none of that context — the gap in answer quality is not small.

These plugins exist to close that gap. They aren't just prompt collections. They are complete role configuration bundles: domain knowledge (Skills) + operational commands (Commands) + tool connections (Connectors via MCP) + specialized sub-agents (Sub-agents).

Author / Team

  • Publisher: Anthropic (the company that builds Claude)
  • Positioning: Official reference implementations and customizable starting points
  • Design intent: Both ready-to-use and templates for teams to build their own plugins

Anthropic's decision to open-source these plugins carries a larger bet: if enough teams start sharing and iterating their own plugins, Claude's expertise in every functional domain will grow in a community-driven way — rather than relying on Anthropic alone to maintain them.

Project Data

  • ⭐ GitHub Stars: 15,500+
  • 🍴 Forks: 1,900+
  • 📄 License: Apache-2.0
  • 🔧 Primary Languages: Python (76.1%), HTML (23.9%)
  • 🔌 Integration Protocol: MCP (Model Context Protocol)
  • 🌐 Repository: anthropics/knowledge-work-plugins

Main Features

Core Utility

Knowledge Work Plugins covers 11 core knowledge work roles:

General Productivity
  └── productivity       Tasks, calendars, daily workflows

Business Functions
  ├── sales              Prospect research, pipeline review, outreach drafting
  ├── customer-support   Ticket triage, responses, KB article creation
  ├── product-management Specs, roadmaps, user research synthesis
  ├── marketing          Content, campaigns, brand voice, performance reporting
  ├── legal              Contract review, compliance checks, risk assessment
  └── finance            Journal entries, reconciliation, financial statements

Technical & Research
  ├── data               SQL querying, visualization, statistical analysis
  ├── enterprise-search  Unified search across email, chat, docs, wikis
  └── bio-research       Life sciences R&D, genomics, literature search

Plugin Tooling
  └── cowork-plugin-management  Create and customize organization-specific plugins
Enter fullscreen mode Exit fullscreen mode

Quick Start

Claude Code Installation:

# Add the plugin source
claude plugin marketplace add anthropics/knowledge-work-plugins

# Install a specific plugin
claude plugin install sales@knowledge-work-plugins

# Install multiple plugins
claude plugin install finance@knowledge-work-plugins
claude plugin install data@knowledge-work-plugins
Enter fullscreen mode Exit fullscreen mode

Claude Cowork Installation:

Install directly from claude.com/plugins — no command line required.

After Installation:

# Skills activate automatically when relevant context is detected
# Slash Commands become available immediately in your session

/sales:call-prep          # Pull CRM data, generate call prep brief
/finance:reconciliation   # Start account reconciliation workflow
/data:write-query         # Enter SQL query assistance mode
/legal:review-contract    # Start contract review
/product:write-prd        # Generate a product requirements document
Enter fullscreen mode Exit fullscreen mode

Deep Dive

The Four-Layer Plugin Architecture

Every plugin follows a consistent directory structure with four layers:

plugin-name/
├── .claude-plugin/
│   └── plugin.json        # Manifest: name, version, permission declarations
├── .mcp.json              # Connectors: MCP server config for external tools
├── commands/              # Commands: explicit slash commands
│   ├── call-prep.md       # Full prompt logic for /sales:call-prep
│   ├── pipeline-review.md # Full prompt logic for /sales:pipeline-review
│   └── ...
└── skills/                # Skills: domain knowledge, auto-activated
    ├── crm-context.md     # CRM data interpretation capability
    ├── objection-handling.md  # Objection handling framework
    └── ...
Enter fullscreen mode Exit fullscreen mode

Each layer has a distinct role:

Layer Trigger Purpose
Skills Auto-triggered (context-based) Domain knowledge Claude draws on when it detects relevance
Commands Explicit (/plugin:command) Specific workflows the user actively initiates
Connectors Called by Skills/Commands Connect to real tools (CRMs, databases, design tools)
Sub-agents Orchestrated by Commands Specialist roles in complex multi-step tasks

Fully Declarative Design: The Power of Zero Code

This is the most important design philosophy in the entire project.

Traditional AI tool integration requires:

  1. Writing API call code
  2. Handling authentication and errors
  3. Managing dependencies and versions
  4. Engineering resources to maintain everything

Knowledge Work Plugins' approach:

<!-- skills/deal-qualification.md -->

## MEDDIC Qualification Framework

When a user asks about a deal or prospect, evaluate using MEDDIC:
- **Metrics**: What quantifiable impact will this solution deliver?
- **Economic Buyer**: Who controls the budget? Have we spoken to them?
- **Decision Criteria**: What criteria will be used to evaluate solutions?
- **Decision Process**: What are the steps to get a deal done?
- **Identify Pain**: What's the specific business problem?
- **Champion**: Who inside the account is advocating for us?

Always surface missing MEDDIC elements as deal risks.
Enter fullscreen mode Exit fullscreen mode

This is a real Skill file. The most complex logic in the entire Sales plugin is expressed as Markdown like this.

What does that mean? Any knowledge worker who understands their own workflow can build or customize a plugin without writing a single line of code. A sales director can encode their sales methodology. A legal team can add their compliance checklist. A data team can document their SQL conventions and data source schemas.

MCP Connectors: The Bridge Between Plugins and the Real World

Knowledge Work Plugins uses MCP (Model Context Protocol) to connect external tools — currently the most standardized AI tool integration protocol.

// .mcp.json (Sales plugin connector config example)
{
  "mcpServers": {
    "hubspot": {
      "command": "npx",
      "args": ["@mcp/hubspot-server"],
      "env": {
        "HUBSPOT_API_KEY": "${HUBSPOT_API_KEY}"
      }
    },
    "clay": {
      "command": "npx",
      "args": ["@mcp/clay-server"],
      "env": {
        "CLAY_API_KEY": "${CLAY_API_KEY}"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

The key design value: connectors and plugin logic are completely decoupled. You can swap the Sales plugin's HubSpot for Salesforce by editing only .mcp.json — every Skill and Command stays untouched.

The 11 Plugins: A Closer Look

Sales Plugin (key connectors: HubSpot, Clay, ZoomInfo, Fireflies):

/sales:call-prep "Acme Corp"   # Pull CRM data, generate pre-call brief
/sales:pipeline-review         # Audit current pipeline, flag at-risk deals
/sales:draft-outreach          # Draft personalized outreach from prospect data
Enter fullscreen mode Exit fullscreen mode

Built-in Skills include: MEDDIC deal qualification framework, objection handling patterns, competitive comparison logic.

Finance Plugin (key connectors: Snowflake, Databricks, BigQuery):

/finance:reconciliation        # Start reconciliation, auto-flag discrepancies
/finance:journal-entry         # Draft journal entries
/finance:close-checklist       # Month/quarter-end close checklist
Enter fullscreen mode Exit fullscreen mode

Built-in Skills include: accounting standard understanding, financial reporting norms, internal control verification patterns.

Data Plugin (key connectors: Snowflake, Databricks, Definite, Hex):

/data:write-query              # Enter SQL query assistance mode
/data:explain-query            # Explain an existing SQL query's logic
/data:build-dashboard          # Guided dashboard design workflow
Enter fullscreen mode Exit fullscreen mode

Built-in Skills include: SQL best practices, data quality check patterns, statistical interpretation frameworks.

Bio-Research Plugin (key connectors: PubMed, ChEMBL, Benchling, bioRxiv):

The most vertical plugin in the library, designed for life sciences R&D:

/bio:literature-search         # Cross-search PubMed + bioRxiv
/bio:compound-lookup           # Query compound data (ChEMBL)
/bio:protocol-draft            # Draft experimental protocols
Enter fullscreen mode Exit fullscreen mode

Legal Plugin (key connectors: Box, Egnyte, Microsoft 365):

/legal:review-contract         # Review contract clauses, flag risks
/legal:draft-nda               # Draft non-disclosure agreement
/legal:compliance-check        # Run compliance self-assessment
Enter fullscreen mode Exit fullscreen mode

Customization: Where the Real Power Is

The official plugins are generic starting points. The real value comes from adapting them to your team.

Typical customization pattern:

Step 1: Add company-specific knowledge to skills/
  → Product terminology glossary
  → Internal process documentation
  → Pricing rules and exceptions

Step 2: Edit .mcp.json to use your actual tools
  → Swap HubSpot for Salesforce
  → Point to your own data warehouse

Step 3: Add company-specific workflows to commands/
  → Your quarterly business review format
  → Your specific approval flows and report templates

Step 4: Share across the team
  → Everyone installs the same plugin = shared expertise
Enter fullscreen mode Exit fullscreen mode

This is the larger vision Anthropic has embedded in this project: as teams build and share their own plugins, Claude becomes an expert in each company and each function specifically — not a generic assistant that has to be re-educated every conversation.

Project Links & Resources

Official Resources

Target Audience

  • Knowledge workers: Sales, legal, finance, product, marketing professionals who want Claude to genuinely understand their job
  • Team leads: Building a consistent, shared AI work pattern across a team
  • AI tool developers: Studying declarative approaches to building domain-specialist plugins
  • Enterprise IT / AI teams: Evaluating how to deploy functional AI tools at organizational scale

Summary

Key Takeaways

  1. 11 functional plugins: Full-spectrum coverage from sales to bio-research, each with real tool connections
  2. Fully declarative: Pure Markdown + JSON, zero code — anyone who knows their work can build and customize
  3. MCP connectors: Standardized tool integration; swap a connector without touching any logic
  4. Four-layer architecture: Skills (auto) + Commands (explicit) + Connectors (tools) + Sub-agents (orchestration)
  5. From Anthropic directly: Represents Anthropic's systematic thinking on how AI actually integrates into knowledge work

One-Line Review

Knowledge Work Plugins isn't just a bundle of prompts — it's Anthropic's answer to the question of enterprise AI adoption: not smarter general models, but better-configured specialists who already know your job.

Find more useful knowledge and interesting products on my Homepage

Top comments (0)