wshobson/agents

What it solves

This project provides a production-ready marketplace of agentic building blocks—including plugins, agents, skills, and commands—that can be used across multiple AI coding assistants. It eliminates the need to rewrite configurations for different tools, allowing a single source of truth for agentic workflows.

How it works

The system uses a single Markdown source to generate harness-native artifacts for five different AI harnesses: Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot. Each plugin is isolated and composable, meaning only the components of an installed plugin are loaded into the AI's context to save resources.

It employs a tiered model strategy to assign different LLMs (such as Opus, Sonnet, and Haiku) to tasks based on the complexity of the work, ranging from long-horizon autonomous tasks to fast operational duties.

Who it’s for

Developers and teams using agentic AI coding assistants who want to access a pre-built library of domain-expert agents (e.g., architecture, security, ML) and specialized skills without manual configuration for each tool.

Highlights

• Multi-Harness Support: Native integration with Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot.
• Extensive Library: Includes 94 plugins, 203 agents, 175 skills, and 109 commands.
• Quality Evaluation: A three-layer framework (plugin-eval) that uses static analysis, LLM judges, and Monte Carlo simulations to certify plugin quality.
• Tiered Model Strategy: Optimizes cost and performance by mapping tasks to the most appropriate model tier.
• External Memory: Integration with Pensyve for external memory capabilities across supported harnesses.