DEV.co
MCP Servers · wshobson

agents

A multi-harness agentic plugin marketplace (90 plugins, 199 agents, 161 skills) designed to run on Claude Code, Cursor, Codex CLI, OpenCode, GitHub Copilot, and Gemini CLI from a single Markdown source. Provides orchestrated AI agents for software development tasks across languages, infrastructure, security, and ML domains.

Source: GitHub — github.com/wshobson/agents
37.6k
GitHub stars
4k
Forks
Python
Primary language
MIT
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Repositorywshobson/agents
Ownerwshobson
Primary languagePython
LicenseMIT — OSI-approved
Stars37.6k
Forks4k
Open issues4
Latest releaseUnknown
Last updated2026-07-07
Sourcehttps://github.com/wshobson/agents

What agents is

Python-based plugin system with auto-discovered agents, skills, and slash commands organized by domain. Generates harness-native artifacts (not lowest-common-denominator translations) via tiered Makefile targets. Includes a three-layer evaluation framework (static, LLM judge, Monte Carlo) and progressive disclosure skill loading to minimize context bloat.

Quickstart

Get the agents source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/wshobson/agents.gitcd agents# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Multi-tool AI agent orchestration

Teams using multiple agentic tools (Claude Code + Cursor + Codex) can manage a single source of truth for plugins, skills, and agents—each tool gets idiomatic, native artifacts without manual translation.

Domain-expert agent scaffolding

Organizations need pre-built, categorized agents (Python, ML, security, infra, architecture) that can be installed granularly. The tiered model strategy (Fable 5 for long-horizon work, Sonnet for fast tasks) helps balance cost and capability.

Cross-platform agentic skill sharing

Development teams can share reusable, certified skills across Anthropic and Google ecosystems without rewriting for each platform. The PluginEval framework (static, LLM judge, Monte Carlo) enables quality gates before adoption.

Implementation considerations

  • Multi-harness generation pipeline requires `make generate` discipline and version control of committed `.codex-plugin/`, `.cursor-plugin/`, `.copilot/` artifacts; drift detection via `make garden` is critical for maintaining parity.
  • Tiered model strategy couples plugin behavior to Anthropic model availability and pricing (Fable 5 for long-horizon tasks); fallback logic and cost controls must be defined in consuming applications.
  • Plugin discovery is directory-structure-based; enforcing strict naming conventions and auto-discovery rules across 90+ plugins requires governance and CI validation (e.g., `plugin-eval score` in pre-commit hooks).
  • Cross-harness compatibility is aspirational; each harness may have idiosyncratic limits (e.g., Codex CLI's 8 KB skill cap), requiring per-harness testing and potential skill fragmentation.
  • External integrations (Pensyve via git-subdir, Gemini upstream repo) introduce transitive maintenance burden; upstream changes in Pensyve may require re-generation and re-validation of all downstream harnesses.

When to avoid it — and what to weigh

  • Single-tool-only requirement — If your team uses only one agentic harness (e.g., Claude Code alone), the multi-harness overhead and Makefile-based generation pipeline add complexity without clear ROI. Simpler, single-format plugin repos may be faster.
  • No model flexibility or budget constraints — The tiered model strategy assumes runtime access to Fable 5 (premium cost), Opus, and Sonnet. If your deployment is locked to a single model or has strict cost caps, the tier-based task routing may not align with your constraints.
  • Closed-source, proprietary workflow requirement — MIT license permits commercial use, but the model is source-visible and extensible. If you need opaque, white-label agent packaging, this marketplace design may expose IP or require careful vendoring.
  • Minimal or no AI agent experience — The marketplace assumes familiarity with prompt engineering, agent design patterns, and multi-model orchestration. Teams new to agentic workflows may struggle with configuration and skill authoring without significant onboarding.

License & commercial use

MIT License. Permissive, OSI-approved; permits commercial use, modification, and distribution with attribution.

MIT license explicitly permits commercial use without restriction. However, the marketplace bundles plugins that may depend on external services (Anthropic, Google, OpenAI models) with their own terms. Consuming applications must review model provider ToS and ensure compliance with any proprietary plugins or integrations (e.g., Pensyve). Source-visible design may require careful IP handling in white-label or closed-source scenarios.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

No explicit security audit, threat model, or signed releases mentioned. Source-visible plugin design allows peer review but also exposes agent prompts and skill logic. Integration with external memory (Pensyve) and multi-model routing (Fable 5, Opus, Sonnet) introduce attack surface (prompt injection, model confusion, cost amplification). Token/rate-limit controls, API key isolation, and model sandboxing are not documented. Organizations should review skill prompts for injection vulnerabilities and implement runtime guardrails (e.g., tool allowlists, token budgets) in consuming applications.

Alternatives to consider

OpenAI Plugins / GPT Store

Simpler, single-harness model; mature ecosystem; lower learning curve. Lacks multi-tool orchestration and tiered model strategy; platform lock-in to OpenAI.

LangChain / LlamaIndex agent frameworks

Language-agnostic, programmatic agent building; larger community. Requires custom skill/plugin abstraction and multi-tool translation; not a pre-built marketplace.

Anthropic's native Claude Code plugin system (alone)

Tightly integrated, low friction for Claude-only teams. No multi-harness support; smaller ecosystem; requires custom solutions for Cursor/Codex/Gemini integration.

Software development agency

Build on agents with DEV.co software developers

Start with Claude Code (`/plugin marketplace add wshobson/agents`), then expand to Cursor, Codex, or Gemini. Review docs/harnesses.md for your target harness setup, capability deltas, and gotchas.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

agents FAQ

Can I use this marketplace with just Claude Code, or do I need all five harnesses?
Yes, Claude Code alone is fully supported. The multi-harness design is optional; if you're single-harness, the marketplace works like any native plugin repo. Multi-harness generation overhead is avoided.
What is the PluginEval framework, and do I need to run it before deploying a skill?
PluginEval is a three-layer evaluation system (static, LLM judge, Monte Carlo) that measures and certifies plugin/skill quality. It's recommended for marketplace quality gates but not mandatory for local use. `uv run plugin-eval certify` provides probabilistic reliability estimates over 50–100 runs.
How do I add my own plugins or skills to this marketplace?
See docs/authoring.md for the portable-content style guide. Plugins are directory-structured (`plugins/<name>/agents/`, `commands/`, `skills/`); agents, commands, and skills are auto-discovered. After authoring, validate with `make validate` and `make garden`, then open a PR.
What if Pensyve (the external memory integration) goes offline or changes its API?
Pensyve is a git-subdir dependency. If upstream breaks, re-generation will fail or produce invalid artifacts. Teams should fork Pensyve or maintain a mirror, and pin to a specific commit hash to isolate from upstream churn.

Work with a software development agency

DEV.co helps companies turn open-source tools like agents into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your mcp servers stack.

Ready to unify your agentic tools?

Start with Claude Code (`/plugin marketplace add wshobson/agents`), then expand to Cursor, Codex, or Gemini. Review docs/harnesses.md for your target harness setup, capability deltas, and gotchas.