Heads up: this post contains affiliate links. If you buy through them we may earn a commission at no extra cost to you — it's how we fund independent research. Full disclosure.
The tool layer for AI agents is fragmenting. Model Context Protocol (MCP), the SKILL.md open standard, ClawHub, and the Hermes Skills Hub are all competing for the same real estate — the mechanism by which agents discover and use capabilities beyond what is baked into the model. Understanding the difference between them, and where each is winning, determines how you invest time in building and publishing skills.
This guide maps every registry and marketplace, explains both competing standards, and explains what a skill written for one platform can — and cannot — do on another.
The two layers: MCP vs. SKILL.md
These are solving adjacent but distinct problems.
Model Context Protocol (MCP) is a server protocol — it defines how an agent runtime connects to an external tool or data source at execution time. An MCP server exposes tools (functions the agent can call), resources (data the agent can read), and prompts (reusable templates). The agent runtime (Claude Desktop, any SDK with MCP support) sends requests to the MCP server; the server executes them and returns results. MCP is about runtime tool connectivity.
SKILL.md is a documentation standard — it defines how an agent’s capabilities are described in a portable, machine-readable format. A SKILL.md file tells any compatible agent what a skill does, how to use it, what tools it requires, and when to invoke it. SKILL.md is about capability discovery and portability.
The two are complementary: an MCP server can be described by a SKILL.md file; a SKILL.md file often references MCP tools. The confusion arises because both are called “skills” colloquially.
The SKILL.md open standard
Anthropic published the SKILL.md open standard (agentskills.io) on December 18, 2025. As of July 2026, it is supported by 40+ agent products including Claude Code, Claude.ai, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI, Goose, and Snowflake Cortex Code.
The cross-vendor adoption is the headline fact here. A SKILL.md file that describes a research workflow can be placed in any repository and picked up by Claude Code, Cursor, and GitHub Copilot without modification. Enterprise integrators publicly documented on agentskills.io include Elastic, Databricks, and Snowflake.
SkillsBench (2026) analyzed 47,150 public skills and found an average quality score of 6.2 out of 12. Curated skills raised agent task pass rates by an average of 16.2 percentage points over uncurated community skills. The quality gap is meaningful — a well-written SKILL.md file is genuinely more valuable than a hastily written one, not just marginally.
The SKILL.md format is open and documented at agentskills.io. The community Discord (discord.gg/MKPE9g8aUy) and the GitHub repository (github.com/agentskills/agentskills) are the primary development venues.
The MCP ecosystem: five registries
| Registry | Server count | Key feature | URL | |
|---|---|---|---|---|
| Official MCP Registry | 200+ enterprise-managed | Namespace authentication via verified GitHub accounts | registry.modelcontextprotocol.io | |
| Smithery | 7,000+ | One-click install, OAuth management | smithery.ai | |
| MCP.so | 19,000+ | Community submissions, largest catalogue | mcp.so | |
| MCP Market | 10,000+ | Browsable web UI, 23 categories | mcpmarket.com | |
| Glama | Partner-maintained | Meta-registry, curated partner list | glama.ai |
The official MCP Registry (registry.modelcontextprotocol.io) uses namespace authentication tied to verified GitHub accounts. A 2026-07-28 specification release candidate was published as of early July 2026. This is the authoritative source for the MCP specification itself, though its server catalogue is smaller than community registries.
MCP.so’s 19,000+ server count is the largest single-catalogue number — but it is community-submitted and unevaluated. Smithery’s 7,000+ servers include OAuth management and one-click installation, which significantly reduces setup friction for common integrations. For a first build, Smithery is the best starting point; for exhaustive search, MCP.so covers more ground.
The Claude Agent SDK’s MCP integration uses a single config block in ClaudeAgentOptions. The Playwright MCP example from the official cookbook: npx @playwright/mcp@latest as the command gives agents browser automation without writing custom tool code. The MCP ecosystem has crossed 200 server implementations in the enterprise-managed tier as of mid-2026.
The skills marketplaces: Hermes Skills Hub and ClawHub
Beyond the MCP registries, two agent-specific skills marketplaces have emerged.
Hermes Skills Hub (Nous Research) hosts 90,000+ community-contributed skills as of July 2026. These are SKILL.md-format skills built specifically for the Hermes Agent runtime. The Awesome Hermes Agent repository (github.com/0xNyk/awesome-hermes-agent) is a curated sub-index of the highest-quality skills.
ClawHub (OpenClaw Foundation) is a centralized plugin registry launched as a one-click install marketplace for the OpenClaw agent. It replaced the previous OpenClaw skills system. ClawHub plugins extend OpenClaw’s OS-level capabilities — additional messaging platform integrations, domain-specific workflows, and tool connections.
The two marketplaces serve different agent runtimes and are not interchangeable. A Hermes skill will not run in OpenClaw without adaptation. However, if the underlying skill is written in SKILL.md format and does not depend on runtime-specific APIs, porting between them is primarily a documentation exercise.
Writing a portable SKILL.md file
The portability value of SKILL.md comes from following the standard faithfully. A skill file that works across 40+ agent products follows this structure:
# Skill: [Name]
## Description
[What this skill does in one sentence]
## When to use
[Specific trigger conditions — not vague "use this for X" but "invoke when the user asks for..."]
## Tools required
[List MCP tools or other dependencies]
## Steps
1. [First step with specific action]
2. [Second step]
## Examples
[One or two concrete input/output examples]
## Limitations
[What this skill does not do well]The SkillsBench quality analysis found that the most common failures in low-scoring skills are: vague “when to use” conditions (agent invokes the skill at wrong times), missing limitation documentation (agent attempts tasks outside the skill’s scope), and absent examples (agent cannot infer correct input format).
Where each standard is winning
SKILL.md is winning the cross-vendor portability battle. 40+ products supporting a standard published in December 2025 is unusually fast adoption for a developer tooling standard. The enterprise integrators (Elastic, Databricks, Snowflake) signal that SKILL.md will remain relevant in production environments.
MCP is winning the runtime connectivity battle. The official registry, Smithery, and MCP.so together give developers access to thousands of tool integrations without writing custom connectors. The Claude Agent SDK’s single-config MCP connection pattern is the cleanest implementation in any framework.
What works
- SKILL.md cross-vendor adoption (40+ products by July 2026) is genuinely fast for a developer standard
- MCP.so's 19,000+ servers mean most common tool integrations already have a ready-made implementation
- Smithery's one-click OAuth management removes the most common setup friction for first-time MCP builders
- The quality gap between curated and uncurated skills is measurable — writing good SKILL.md files is worth the effort
- Enterprise adoption (Elastic, Databricks, Snowflake) signals SKILL.md will be maintained and extended
What doesn't
- Hermes Skills Hub and ClawHub are not interchangeable — skills built for one runtime require adaptation for another
- MCP.so's 19,000 servers are community-submitted and unevaluated — quality varies significantly
- The 2026-07-28 specification release candidate for the official MCP Registry means the standard is still evolving
- Cross-vendor portability is only as good as each product's SKILL.md implementation — some integrators may not fully follow the standard
Common questions
Do I need to write separate skills for Claude Code, Cursor, and GitHub Copilot?
No. One SKILL.md file placed in a repository will be read by all three products (and 37+ others supporting the standard) without modification. The standard was specifically designed for this portability. Where products diverge is in how they invoke skills — the trigger conditions in your SKILL.md description affect how confidently each agent recognizes when to use the skill.
What is the difference between an MCP server and a SKILL.md file?
An MCP server is running infrastructure — a process that receives requests from an agent and executes them. A SKILL.md file is documentation — a text file that describes what capabilities an agent has and how to use them. An MCP server tells the agent how to call a tool; a SKILL.md file tells the agent when to use it and how to compose it into a workflow.
How do I find the best MCP servers without evaluating all 19,000?
Start with Smithery for curated servers with one-click install. For specific integrations, search by category on MCP Market (10,000+ servers across 23 categories). The official MCP Registry is the right source for enterprise-grade, namespace-authenticated servers. MCP.so is the right source for breadth — use it when you cannot find what you need elsewhere.
Should I publish skills I build to the Hermes Skills Hub or to a SKILL.md standard repository?
Both if you can. Publishing to the Hermes Skills Hub gives you access to a specific community of 210,000+ Hermes users. Publishing as a SKILL.md file in a public GitHub repository makes it discoverable by any agent product supporting the standard — broader reach, less targeted audience. For skills with general applicability, SKILL.md is the higher-leverage publication venue.