AI Skills & MCP Technical Glossary

An open standard that enables AI assistants to securely connect to external tools, databases, and APIs.

Model Context Protocol (MCP) is an open-source standard pioneered by Anthropic that normalizes how frontier models interact with local environments and remote resources. MCP separates concerns between clients (Cursor, Claude Desktop, Windsurf) and tool providers (PostgreSQL, GitHub, Slack) via standard JSON-RPC 2.0 communication over stdio or SSE transport layers.

A standardized markdown format defining executable instructions, tools, and prerequisites for AI agents.

SKILL.md is an open convention for packaging AI directives and agent capabilities inside repositories. It features YAML frontmatter (defining name, description, required CLI tools, and environment variables) followed by markdown instructions that AI coding assistants like Claude Code and Codex CLI ingest as high-level operational directives.

An agentic command-line interface from Anthropic that executes coding workflows directly in terminal environments.

Claude Code is an autonomous coding agent operating inside the user's terminal. It reads file directory trees, searches codebases, edits multi-file modules, runs automated tests, and creates git commits autonomously based on user prompts.

The two primary communication channels used by Model Context Protocol servers.

Stdio transport launches the MCP server as a local child process and communicates via standard input/output streams, offering zero-latency execution without network port collisions. SSE (Server-Sent Events) transport runs the MCP server as an HTTP service, allowing remote hosted integrations and cloud multi-tenant execution.

Abstract Syntax Tree static analysis verifying that AI-generated code conforms to security constraints.

AST Sandboxing parses source code into an abstract syntax tree prior to execution to detect unsafe system calls, arbitrary shell injections, or unconstrained filesystem access, ensuring AI agents operate within defined security boundaries.

A directed acyclic graph where each AI execution step has verifiable inputs, error handlers, and predictable outputs.

Unlike unpredictable single-prompt loops, deterministic DAG workflows break complex tasks into explicit state machines where intermediate steps are validated with schemas and automated assertions, yielding 99.9% reproducible outcomes in production pipelines.

A security vulnerability where untrusted input tricks an AI model into overriding its system directives.

Prompt injection occurs when user data or third-party web content contains natural language directives designed to hijack the model's control flow. AIStoreHub uses automated heuristic scanners to detect and neutralize prompt injection vectors in indexed packages.

JSON Schema specifications that define the exact arguments an AI model can pass to executable functions.

Function calling enables models to output structured JSON matching a predefined JSON schema instead of freeform prose. The runtime executes the underlying tool with those typed arguments and returns the result back to the model context.