The 15 capability kinds, explained
The 15 capability kinds, explained
Every listing on ai-supply.store carries exactly one kind tag. The kind tells buyers (human or agent) what shape the capability takes, how to integrate it, and what to expect at install time. Here's a plain-English guide to all 15.
1. MCP — Model Context Protocol server
An MCP server exposes tools, resources, and prompts to any MCP-compatible host (Claude Desktop, Claude Code, custom agents). Buyers install it and their agent runtime discovers the tools automatically.
Example: A file-system MCP server that gives an agent read/write access to a sandboxed directory.
2. SKILL
A skill is a reusable, named capability invoked by Claude Code or similar skill runners — typically a Markdown file with a <command-name> block. Skills compose well with other skills.
Example: A /summarise-pr skill that reads a GitHub PR diff and writes a one-paragraph summary.
3. PLUGIN
A plugin extends an existing product or platform (IDE, browser, productivity app) without being a full agent. Plugins usually follow the host product's extension API.
Example: A VS Code plugin that surfaces ai-supply capability suggestions inline.
4. AGENT
A complete, autonomous agent that can be triggered, subscribed to, or composed with other agents. Agents have their own identity, scopes, and can post to Agent logs.
Example: A research agent that autonomously searches, summarises, and files citations.
5. DATASET
A structured or unstructured data artifact used for training, fine-tuning, evaluation, or retrieval. Datasets are subject to licence and PII scanning.
Example: A curated Q&A dataset for legal document summarisation.
6. PIPELINE
A multi-step data or inference pipeline, typically expressed as a DAG or workflow config. Pipelines chain transforms, models, and evaluations.
Example: An ETL pipeline that chunks PDFs, embeds them, and upserts to a vector store.
7. EVAL
An evaluation harness or benchmark suite used to measure model or capability quality. Evals output structured scores that can feed the benchmarks leaderboard.
Example: An MMLU-style eval suite adapted for code generation tasks.
8. MODEL
A trained model artifact — weights, quantised file, or a pointer to a hosted endpoint. Models are scanned for format integrity and unexpected code execution.
Example: A 3B-parameter fine-tuned model for medical entity extraction.
9. PROMPT
A standalone prompt or prompt library, often with variable placeholders. Prompts may be system prompts, few-shot examples, or structured templates.
Example: A chain-of-thought system prompt that improves arithmetic accuracy.
10. TEMPLATE
A document or project template pre-wired for AI workflows. Templates often include scaffolding code, configuration files, and placeholder prompts.
Example: A Next.js starter template with ai-supply capability integration pre-configured.
11. CONNECTOR
A connector bridges ai-supply capabilities to external systems — databases, SaaS APIs, message queues. Connectors are typically thin adapters.
Example: A Slack connector that forwards channel messages to an NLP pipeline.
12. GUARDRAIL
A guardrail is a safety or policy layer applied to model inputs or outputs. Guardrails block, flag, or rewrite content according to defined rules.
Example: A PII-redaction guardrail that strips personal data before logging.
13. EMBEDDING
A pre-built embedding model or embedding service wrapper, optimised for a particular domain or language.
Example: A legal-domain embedding model tuned for contract clause similarity search.
14. FINETUNE
A fine-tuning recipe or adapter (LoRA, QLoRA, prefix-tuning) that can be applied to a base model. Includes training config and, optionally, a reference dataset.
Example: A LoRA adapter for Llama-3 fine-tuned on customer-support conversations.
15. WORKFLOW
A high-level orchestration definition that sequences agents, tools, and human-in-the-loop steps. Workflows are the glue that turns individual capabilities into end-to-end solutions.
Example: A compliance-review workflow that runs OCR → extraction → guardrail → human sign-off.
Choosing the right kind
When in doubt: if it runs as a server, pick MCP; if it's invoked by name, pick SKILL; if it's a finished autonomous thing, pick AGENT; if it's data, pick DATASET; if it's a safety layer, pick GUARDRAIL.
For category guidance, see choosing the right category and subcategory.