Definition
An MCP server is a program that exposes tools, resources and prompts to AI applications over the Model Context Protocol, either as a local process speaking over standard input and output or as a remote service over HTTP.
Also called: Model Context Protocol server, MCP tool server.
The Model Context Protocol is the standard; an MCP server is one concrete program that speaks it. The AI application (the host, such as Claude Code or Codex) starts or connects to the server, asks what it offers, and lets the model call its tools during a chat.
| Local server | Remote server | |
|---|---|---|
| Runs | as a process on your machine, started by the host | as a web service |
| Transport | stdio (standard input and output) | HTTP |
| Typical use | files, local databases, command-line tools, apps on your Mac | hosted services such as issue trackers or design tools |
| Access | whatever your account can reach | whatever the account you sign in with can reach |
Claude Code adds servers with claude mcp add, a project's .mcp.json, or through a plugin; Codex lists them in config.toml or gets them from a plugin. Each server's tool definitions take up some of the model's context, and a tool acts with the server's access, so a server that only reads is the safer kind to allow broadly.
In Pairbar
Pairbar ships its own local MCP server, pairbar mcp, installed as a plugin for Claude and Codex (Settings → Integrations → Pairbar tools, off until you install it). Most of its 22 tools only read, such as list_chats, list_tasks, list_skills and get_limits, limited to the calling chat's project by default and never returning another chat's messages. The others are agent actions such as second_opinion or run_workflow, which run only if your settings allow them or you approve them in Pairbar. In Claude it shows in /mcp as plugin:pairbar:pairbar. Every call is logged in Agent activity. See let Claude and Codex query Pairbar.
Frequently asked questions
- What's the difference between MCP and an MCP server?
- MCP is the protocol, the shared language. An MCP server is a program that speaks it to offer tools and data; the host app (Claude Code, Codex) is its client.
- Is a local MCP server safe?
- It runs with your account's access, so trust depends on what its tools do. Read-only servers are lower risk; still add only servers you trust.
- Can one MCP server work in both Claude Code and Codex?
- Yes. Both speak MCP, though each is configured separately. Pairbar's
pairbar mcp, for example, serves both from one plugin folder.
Related
- GlossaryModel Context Protocol (MCP)An open standard for connecting AI applications such as Claude Code and Codex to external tools, data sources and workflows.
- GuideLet Claude and Codex query Pairbar (MCP tools)Install Pairbar's plugin so Claude and Codex chats can call its 22 tools (MCP or the pairbar command): read what Pairbar knows, ask for approved actions.
- DocsIntegrationsLive status for Claude and Codex through their hooks, Pairbar's tools for their chats (the plugin, the `pairbar` command, the tool reference, who's calling, agent activity), and agent actions: second opinions, reviews, helper chats, handoffs, workflow runs and drafts, and check-ins that agents ask for and you approve, with budgets and guardrails
- GlossaryAgent hooksCommands an AI coding agent runs automatically at set points in its loop, such as before a tool call, on a permission request or when a turn stops.
- GlossaryAgent actionSomething an AI coding agent asks a tool to do on its behalf (start a chat, run a job), as opposed to only reading information.