MCP vs RAG vs Agent Skills vs Custom GPTs: A Decision Tree for Companies

These four are not four answers to the same question. They sit at different layers of an AI system. RAG is a retrieval pattern for supplying relevant external context to generation. MCP is an open protocol through which AI applications can discover context and tools exposed by servers. Agent Skills are an open folder format for procedural instructions and optional resources or scripts. Custom GPTs are configured, no-code assistants used inside ChatGPT. A compatible agent can combine retrieval, MCP tools, and Skills, but none of those layers automatically supplies the others. A Custom GPT is a separate ChatGPT product choice with account, workspace, and capability constraints.

This is an engineering view, not a vendor pitch. We checked the product and specification claims against primary sources on 2 September 2026. Product availability and host support can change, so re-check the linked documentation before committing.

The distinction at a glance

The fastest way to stop conflating them is to see what each one actually is.

The first three can be composed in one compatible agent. A Custom GPT is a ChatGPT delivery surface. OpenAI's Apps SDK builds on MCP, but that does not mean every Custom GPT or GPT Action is an MCP integration.

RAG: answering from your knowledge

Retrieval-Augmented Generation combines retrieved external context with generation. It can ground answers in proprietary or current material without retraining, but it does not guarantee factual output: retrieval quality, source quality, prompting, and evaluation still matter. The retrieval stage itself does not write to an API or change state. An application can combine RAG with action-taking tools, but those actions come from another layer. If you are weighing RAG against fine-tuning or a bigger context window, see RAG vs fine-tuning vs long context.

MCP: connecting the model to your systems

The Model Context Protocol is an open client-server standard introduced by Anthropic in November 2024. Under the current specification, servers can expose prompts, resources, and tools; clients and servers negotiate supported capabilities. A connector can therefore be reused by compatible hosts, but deployment, authorization, supported protocol features, and tool approval still need host-specific testing. MCP does not itself define retrieval, a model, or a business procedure. Adoption is demonstrably cross-vendor: OpenAI added remote MCP to the Responses API, Google made Gemini API and SDK tools MCP-compatible, and Microsoft made MCP integration generally available in Copilot Studio. On 9 December 2025, MCP became a founding contribution to the Linux Foundation's Agentic AI Foundation.

Agent Skills: packaging the procedure

Anthropic introduced Agent Skills in October 2025 and published the format as an open standard in December. The specification defines a directory containing a required SKILL.md with YAML frontmatter and Markdown instructions, plus optional scripts, references, and assets. Compatible agents use progressive disclosure to load more detail when needed. Runtime support, allowed tools, dependencies, and compatibility can still differ. Skills can describe how to use MCP tools, but the runtime must provide and authorize those tools. Treat scripts in a skill like installed software and review their source and permissions.

This is not theory for us. We publish a public Agent Skills registry. Its public response lists six skills in a machine-readable manifest at /.well-known/agent-skills/index.json, with a SHA-256 digest for each SKILL.md and a human-readable page. Some skills also include reference files, so they are not all single-file packages. Our qa-advisor skill, for example, packages a QA review procedure and supporting references. The manifest digest lets a consumer check file integrity; it is not a publisher identity signature.

For a concrete writing use case, our guide to making AI writing more natural with Agent Skills compares three open-source approaches and shows how to evaluate a team workflow without erasing the author's voice.

Custom GPTs: the quick packaged assistant

OpenAI describes GPTs as no-code assistants built and used inside ChatGPT. They can combine instructions, uploaded knowledge, built-in capabilities, Apps, or Actions. A GPT can use Apps or Actions, but not both at the same time. The current creation documentation says new GPT creation and publishing are unavailable on personal Free, Go, Plus, and Pro accounts; permitted Business, Enterprise, and Edu workspace users can create them on the web. Existing personal GPTs remain usable and may remain editable subject to plan and permissions. A GPT can hold up to 20 knowledge files, each up to 512 MB, and capability availability depends on account, workspace, and region. GPTs run inside ChatGPT and are not an embedding mechanism for an external product. Use one when that managed ChatGPT surface and its constraints fit the job; use an API application when you need your own product surface, deployment, or programmatic lifecycle.

The decision tree

Start from what you actually need the model to do.

1. Answer from your documents or knowledge? RAG. Ground answers in your data, no retraining.
2. Expose systems, APIs, databases, or internal tools through a reusable protocol? MCP. Build a server, then test authorization and capabilities in every intended host.
3. Reliably follow a repeatable procedure or company workflow? Agent Skills. Package the know-how, load it on demand.
4. Want a no-code assistant inside an eligible managed ChatGPT workspace? Custom GPT, if workspace policy and permissions allow creation.
5. Need more than one of these (most real systems)? Combine them. A support agent answers from docs (RAG), looks up an order and issues a refund (MCP), and follows your escalation and tone rules (Skill). RAG can even be exposed as an MCP tool, so "RAG vs MCP" is usually a false choice.

"RAG is knowledge, MCP is connectivity, Skills are procedure, and a Custom GPT is a packaged assistant. Once you see them as different layers instead of competing products, the question stops being which one and becomes which combination."