Two Ways Claude Code Interacts With the World
When Claude Code takes an action, it uses a tool. Reading a file is a tool call. Running a bash command is a tool call. Fetching the contents of a URL is a tool call. The distinction that confuses most developers is between tools that come built into Claude Code and tools that come from MCP servers.
They are not competing systems. They serve different purposes and can be used in the same session, often in the same task.
Built-In Tool Use: What It Is
Tool use refers to Claude’s ability to call predefined functions to interact with its environment. In Claude Code, these tools ship with the product and are always available without any configuration.
The built-in tools in Claude Code include:
- Bash: executes shell commands in your terminal
- Read: reads file contents from your local filesystem
- Write: creates or overwrites files
- Edit: makes targeted edits to existing files
- Glob: finds files matching a pattern
- Grep: searches file contents with regex
- WebFetch: fetches the contents of a URL
- WebSearch: searches the web via a built-in integration
These tools are defined internally by Anthropic. You cannot add to them or modify them. They are optimized for interacting with your local development environment.
Built-in tools give Claude control over your local machine. MCP tools give Claude reach into external services.
MCP Servers: What They Add
MCP (Model Context Protocol) is an open protocol for connecting Claude to external servers that expose their own tool sets. When you install and configure an MCP server, Claude gains access to a new category of tools that the server defines.
The GitHub MCP server defines tools like list_pull_requests, create_issue, and get_file_contents (from remote repos). The Postgres MCP server defines tools like query, describe_table, and list_schemas. These tools do not exist in the built-in set.
MCP servers run as separate processes on your machine. Claude communicates with them over the MCP protocol. From Claude’s perspective, MCP tools and built-in tools behave identically during a session. It calls them the same way and reasons about their outputs the same way.
The difference is in what they connect to: built-in tools interact with your local environment, while MCP tools interact with external services and systems.
Side-by-Side Comparison
| Dimension | Built-in Tool Use | MCP Servers |
|---|---|---|
| Configuration required | None | Yes (install + config file) |
| What they connect to | Local files, shell, web | External services and APIs |
| Examples | Bash, Read, Write, Grep, WebFetch | GitHub, Supabase, Postgres, Notion |
| Always available | Yes | Only when configured |
| Defined by | Anthropic | Server authors (community or vendors) |
| Write access | Local filesystem and shell | Depends on the server |
| Custom tools possible | No | Yes (build your own MCP server) |
| Startup overhead | None | Small process per server |
When to Use Built-In Tools
Built-in tools are the right choice for anything that lives on your machine or can be reached over the open web.
Use built-in tools when you need to:
- Read, write, or edit files in your project
- Run build commands, tests, or scripts
- Search your codebase for symbols or patterns
- Fetch public documentation from a URL
- Run database migrations locally via CLI
- Interact with services through their command-line interfaces
If the task can be accomplished through the shell or the filesystem, a built-in tool handles it. There is no need for MCP in those cases.
When to Use MCP Servers
MCP servers are the right choice when Claude needs to interact with a service through its API rather than through the command line.
Use MCP servers when you need Claude to:
- Read or write GitHub issues and pull requests through the API
- Query a remote database without exposing a connection string in the shell
- Read Notion pages or databases as structured context
- Create Linear issues from within a session
- Search the web through a specific search engine API
The key distinction is API access vs shell access. If the service has a CLI and you are comfortable running it in Bash, the built-in Bash tool often works fine. If the service is better accessed through structured API calls with typed inputs and outputs, an MCP server provides a cleaner interface.
Can You Use Both in the Same Session?
Yes, and this is common. Most real development sessions use both.
A typical workflow might look like this:
- Claude uses the GitHub MCP server to read the details of an open pull request
- Claude uses the built-in Read tool to read the relevant local files
- Claude uses the built-in Edit tool to make changes based on the PR comments
- Claude uses the built-in Bash tool to run tests
- Claude uses the GitHub MCP server to post a comment on the PR
Steps 1 and 5 require MCP. Steps 2, 3, and 4 use built-in tools. Claude orchestrates both transparently without you needing to specify which to use.
The decision of which tool to call is Claude’s, not yours. Your job is to ensure the right tools are available.
A Note on Custom Tool Use vs Custom MCP Servers
When developers refer to “tool use” in the context of the Claude API (rather than Claude Code), they mean something slightly different: the ability to define custom tools in an API request that Claude can call during a conversation. This is a feature of the Anthropic API used when building Claude-powered applications.
Claude Code’s built-in tools are a fixed set. You cannot add custom tools to Claude Code through the API tool-use mechanism. If you want to add custom tools to Claude Code, you do it by building a custom MCP server. The Claude Code course covers both built-in tool use and MCP configuration in depth. That server can expose any tools you define.
The guide to building a custom MCP server covers this in detail.
Frequently Asked Questions
Does using MCP servers replace any built-in tools?
No. Built-in tools are always available regardless of which MCP servers you have configured. Adding MCP servers only adds to what Claude can do. It does not remove or replace existing capabilities.
Can MCP servers execute shell commands the way the Bash tool does?
Only if the MCP server explicitly exposes a tool for running commands, which is unusual. Most MCP servers expose structured API calls to specific services rather than general shell access. The built-in Bash tool remains the primary way to execute shell commands in Claude Code.
Is MCP more powerful than built-in tool use?
Neither is more powerful in absolute terms. They serve different purposes. Built-in tools give deep, fast access to your local environment. MCP servers give structured, authenticated access to external services. A Claude Code setup that uses both is more capable than one that uses only either.
Do I need to understand the difference to use Claude Code effectively?
Not necessarily. Claude Code manages the distinction internally and calls whichever tools are appropriate. The practical reason to understand the difference is so you know when configuring an MCP server will help versus when the built-in tools already handle the task. If Claude keeps failing at something because it cannot reach an external service, an MCP server is the solution.
Wondering which tools your Claude Code setup is missing?
Audit your current workflow against the built-in tool list, then identify which external services you are context-switching to manually, those are the gaps an MCP server fills.
Path one: explore it yourself. Check which built-in tools already cover your workflow, then look at the list of the best MCP servers to identify what is missing. The Claude Code course walks through how to combine built-in tools and MCP servers in practical workflows.
Path two: work with Phos AI Labs. If you are setting up Claude Code for an engineering team and want the right combination of MCP servers and built-in tool workflows designed from the start, we run that implementation. Phos AI Labs is a CCA-F certified Claude implementation partner. Thirty minutes, no deck. Start here.