AI Agents, Codex, and MCP

Draft B. Drafted with GPT-5.5 Extra High.

Codex is a coding-focused AI agent. MCP is the protocol that lets Codex connect to external tools, documents, data, and services in a standard way.

LLMs and AI Agents Are Different

An LLM is a model that receives input and generates likely next text. It is close to the reasoning and language-generation engine.

An AI agent is an application built around an LLM that can pursue a goal through multiple steps, call tools, maintain state, and act within a defined environment. OpenAI describes agents as systems that act independently on a user’s behalf and are built from models, tools, and instructions. It also distinguishes agentic systems from simple one-shot LLM calls. (openai.com)

A compact distinction is:

LLM = reasoning and language generation capability
AI agent = an application that uses an LLM plus tools and workflow to do work

An agent is not unconstrained autonomy. It acts inside permissions, tools, rules, and approval boundaries.

What Codex Is

Codex is OpenAI’s AI agent for coding work. Codex CLI is a coding agent that runs on the local computer, and Codex can also appear through an IDE extension or Codex Web. The OpenAI Codex repository describes Codex CLI as a coding agent that runs locally. (github.com)

The important point is that Codex is not the LLM itself. Codex includes an LLM, but Codex is a larger application.

LLM
-> understands text, reasons, plans, and generates responses
 
Codex
-> uses an LLM to read codebases, choose edits, call tools,
   and perform software-development tasks

Codex is an agent specialized for coding. The LLM provides much of the reasoning and generation capability inside that agent.

What MCP Is

MCP, or Model Context Protocol, standardizes how AI applications connect to external systems. The official MCP documentation describes MCP as an open standard that lets AI applications connect to external systems such as local files, databases, search tools, calculators, and workflows. (modelcontextprotocol.io)

MCP does not make the AI smarter by itself. It does not reason, plan, or write code. It standardizes how an AI agent can discover tools, call tools, read resources, and use prompts.

MCP commonly separates three roles:

MCP Host
-> an AI application such as Codex, Claude Desktop, or VS Code
 
MCP Client
-> the component inside the host that maintains a connection to one MCP server
 
MCP Server
-> a program that exposes external functionality or data through MCP

The MCP architecture documentation separates host, client, and server, and describes MCP servers as programs that provide context. MCP can expose tools, resources, and prompts. (modelcontextprotocol.io)

The Core Relationship

Codex and MCP relate in two directions.

First, Codex can act as an MCP host or client and use external MCP servers. OpenAI documentation says Codex supports MCP servers in the CLI and IDE extension, and that MCP can let Codex interact with third-party documentation, browsers, Figma, and other development tools. (developers.openai.com)

The flow looks like this:

User
-> asks Codex to inspect a Figma design and build a React component
 
Codex
-> understands and plans the task
 
Codex MCP client
-> connects to a Figma MCP server
 
Figma MCP server
-> exposes design data or tools through MCP
 
Codex
-> writes or modifies code using the returned information

Second, Codex itself can also run as an MCP server. OpenAI documentation describes codex mcp-server, which lets another MCP client or an Agents SDK based agent call Codex. In that setup, Codex can become a specialized coding executor inside a larger multi-agent workflow. (developers.openai.com)

That flow looks like this:

Higher-level AI agent
-> plans the overall work
 
Codex MCP server
-> performs coding-specific implementation or review work
 
Higher-level AI agent
-> consumes the Codex result and continues the workflow

So the relationship is not one-directional:

1. Codex can use MCP servers.
2. Codex can also be exposed as an MCP server.

What MCP Avoids

Without MCP, every AI agent would need a custom integration for every external tool.

For example, if Codex needs Figma, Jira, GitHub, internal docs, databases, and browser automation, then without MCP each integration needs its own connection rules, authentication style, tool description format, and call format.

With MCP, an MCP server wraps the external system and exposes tools in a standard shape. Codex can inspect those tools and call them through the MCP connection. Codex CLI documentation says configured MCP servers are launched at session start and expose MCP tools alongside built-in tools. (developers.openai.com)

MCP is not a Codex-only feature. It is a common connection protocol that many AI agents can use.

Boundaries That Prevent Confusion

MCP is not an AI agent. It does not judge or plan. Codex and the LLM inside it do the judgment.

An MCP server is not necessarily the external system itself. A GitHub MCP server is closer to an adapter that exposes GitHub functionality through MCP.

Codex is not MCP. Codex is a coding agent. MCP is a protocol that lets Codex connect to external tools.

The LLM does not directly call MCP servers. In practice, the host application uses the LLM’s output to route a tool call through an MCP client, then feeds the result back into the model context. MCP architecture documentation describes this pattern: the AI application gathers tools into a registry available to the LLM, and when the model chooses a tool, the application routes the call to the relevant MCP server. (modelcontextprotocol.io)

Simple Mental Model

LLM = reasoning and language generation capability
AI agent = an application that uses an LLM to perform multi-step work
Codex = an AI agent specialized for coding tasks
MCP = a standard protocol for connecting agents to external tools, data, and services
MCP server = an adapter that exposes a specific external capability through MCP

Codex is the worker. MCP is the standard connection layer that lets the worker use more tools.

Core Takeaway

Codex is an AI agent that uses an LLM to perform coding work. MCP is the protocol that lets Codex discover and call external tools, documents, and services through a standard interface. MCP is not Codex’s brain. It is a standard connection layer that expands the environments Codex can work with.