640. Skill API Standards and Specifications
Basic Information
| Item | Details |
|---|---|
| Topic | AI Agent Skill/Tool API Standards and Specifications |
| Core Protocols | MCP (Tool Layer), A2A (Agent-to-Agent), WebMCP (Web Access) |
| Governing Body | Agentic AI Foundation (AAIF), Linux Foundation |
| Co-founders | Anthropic, Block, OpenAI |
| Supporting Companies | 100+ companies, including Google, Microsoft, Amazon |
| Key Standards | Agent Skills Specification, SKILL.md File Format |
Product Description
The AI Agent Skill API Standards and Specifications are the infrastructure ensuring interoperability among different AI agents, tools, and platforms. By 2026, the industry had reached a consensus on a three-layer AI protocol stack:
- MCP (Model Context Protocol): Communication standard between AI agents and tools
- A2A (Agent-to-Agent): Communication standard between AI agents (led by Google, merged with IBM ACP)
- WebMCP: Standard for AI agents accessing web pages
These three layers are governed by the Agentic AI Foundation (AAIF), a directed fund established under the Linux Foundation in December 2025, co-founded by Anthropic, Block, and OpenAI, and supported by over 100 companies including Google, Microsoft, and Amazon.
Detailed Explanation of the Three-Layer Protocol Stack
MCP - Model Context Protocol (Tool Layer)
- Release: November 2024 (Anthropic)
- Donation: December 2025 (Linux Foundation/AAIF)
- Function: Defines how AI agents discover and use external tools
- Core Concepts: Tools (tool invocation), Resources (data access), Prompts (prompt templates)
- Transmission Methods: stdio (local), HTTP+SSE (remote)
- SDK Downloads: 97M+/month (Python+TypeScript)
- Active Servers: 10,000+
- Adopting Vendors: Anthropic, OpenAI, Google, Microsoft, Amazon
A2A - Agent-to-Agent Protocol (Inter-Agent Communication)
- Release: 2025 (led by Google)
- Merger: IBM's ACP protocol merged into A2A in August 2025
- Function: Defines how AI agents discover, communicate, and collaborate
- Core Concepts: Agent Cards (agent descriptions), Task Protocol (task protocol)
- Use Cases: Multi-agent collaboration, cross-organization agent communication
WebMCP (Web Access Layer)
- Function: Standardizes how AI agents access and manipulate web pages
- Status: Proposed in 2026 as part of the three-layer stack
Agent Skills Specification
Overview
In December 2025, Anthropic released the Agent Skills Specification as an open standard, later adopted by OpenAI. Skills operate at a higher abstraction layer than MCP—a skill may orchestrate multiple tool invocations as part of a workflow.
SKILL.md File Format
- Format: Markdown file + supporting files
- Content: Instructions, context, tool invocation guidelines
- Versioning: Semantic versioning (semver)
- Distribution: Published via registries like ClawHub
- Compatibility: Supports 16+ AI tools (Claude Code, Cursor, OpenAI Codex, Gemini CLI, etc.)
Skills vs Plugins vs MCPs
| Dimension | Skills | Plugins | MCPs |
|---|---|---|---|
| Nature | Knowledge/Instructions | Code Packages | Connectors |
| Format | Markdown File | Code Modules | Server Programs |
| Installation | Place in Project Directory | Package Manager Installation | Start Service Process |
| Token Consumption | Minimal | Moderate | High |
| Complexity | Lowest | Moderate | Highest |
| Persistence | Cross-session Persistent | Environment Dependent | Requires Running Service |
Key Technical Standards
Security Standards
- Minimal Permissions: MCP server permissions should be precisely limited to what the agent requires
- Explicit Authorization: Environment variables and MCP connections require explicit user consent
- Chain of Trust: Signing and verification similar to Docker Content Trust
- Security Scanning: Automatic security scanning upon release
- Revocation Mechanism: Rapid revocation upon discovery of malicious skills
Version Management
- Semantic Versioning: Major.Minor.Patch
- Changelog: Records changes for each version
- Backward Compatibility: Minor version upgrades maintain backward compatibility
- Lock Files: Precisely lock dependency versions
Business Model
- Protocol standards are free and open
- Standard compliance certification services
- Enterprise-grade MCP server hosting
- Security audits and compliance consulting
- Skill development tools and IDE plugins
Target Users
- AI Tool Developers: Build tool integrations compliant with standards
- Enterprise Architects: Plan AI agent infrastructure
- Platform Operators: Build skill marketplaces
- Standard Setters: Participate in protocol evolution
- Security Researchers: Study protocol security
Relationship with the OpenClaw Ecosystem
The Skill API Standards and Specifications form the technical foundation of the OpenClaw ecosystem. OpenClaw's skill system is built on these standards:
- MCP provides the tool connection layer (how AI agents invoke external APIs and services)
- Agent Skills provide the skill abstraction layer (higher than MCP, orchestrating multiple tool invocations)
- SKILL.md is the file format standard for skills
- ClawHub is the platform for skill discovery and distribution
The openness and vendor-neutral governance of these standards ensure that OpenClaw is not locked into a single vendor while benefiting from the contributions of the entire industry. The fact that AAIF is supported by over 100 companies demonstrates the industry's recognition of these standards.
Information Sources
External References
Learn more from these authoritative sources: