Code
Active
Claude Code
No plug-and-play model swapping, no third-party backends. Here's exactly what that means, what Claude Code can do, and whether it's the right tool for your workflow.
Claude Code is Anthropic's command-line AI coding agent β and it has a very deliberate restriction baked in from day one: it runs exclusively on Claude models.
If you've used GitHub Copilot or Cursor, you have a rough frame of reference, but Claude Code operates at a different level of abstraction. Rather than completing individual lines or suggesting functions, Claude Code behaves more like a junior developer who lives inside your terminal. You hand it a task in plain English; it reads your files, writes or edits code, runs tests, interprets errors, and iterates β all without you needing to manually copy output between tools.
It's an agentic tool, meaning it takes sequences of actions autonomously. That's a meaningful distinction. Most AI code assistants respond to prompts. Claude Code acts on goals.
This is probably the most common question developers ask when they first encounter Claude Code β and it's a fair one. The short answer is that Claude Code was built from scratch around Claude's specific capabilities, including its extended context window, its reasoning patterns, and its constitutional AI training.
Claude Code relies on Claude's ability to hold massive amounts of context at once β entire codebases, chains of tool calls, multi-step reasoning β without losing track of earlier state. The system prompts and tool-use architecture underpinning Claude Code are tightly coupled to how Claude processes structured inputs and chains thoughts together. Swapping in a different model would break the tool-calling loop that makes the whole thing work.
Claude Code gives you a choice of models depending on your needs β but every option is a Claude model. Here's how the current lineup breaks down:
You can set which model Claude Code uses via the ANTHROPIC_MODEL environment variable or through the --model flag at the command line. For most individual developers, Sonnet is the sweet spot. Opus is worth the extra cost when you're working through something genuinely complex β like migrating a legacy monolith or debugging a subtle concurrency issue across a large codebase.
Rather than a vague list of "AI features," here's a practical breakdown of what Claude Code can handle, based on how it actually works in a real terminal session.
Claude Code doesn't just edit one file at a time. Ask it to rename a function across your entire codebase and it will find every reference, update every call site, and show you a diff before committing anything.
It can run npm test, interpret the output, notice a failing test, trace it back to a specific function, fix it, and re-run the suite β all in one uninterrupted flow.
Point Claude Code at an unfamiliar repo and ask it to explain the architecture. It reads the actual source code β not documentation β and gives you a real mental model of how the system is structured.
Claude Code understands git context. It can review your diff before a commit, write a meaningful commit message based on actual changes, or help you untangle a messy merge conflict with full awareness of both sides.
Give it a module and ask for test coverage. Claude Code writes unit tests that reflect actual edge cases, not just happy paths β because it reads the implementation before it writes the test.
Claude Code supports the Model Context Protocol, letting you connect external data sources β databases, internal APIs, documentation systems β directly into its context. Your company's domain knowledge becomes part of the conversation.
The tool is flexible, but it's not magic. Here are the situations where developers consistently report getting real value out of Claude Code β and where the Anthropic-model architecture actually matters.
01 Legacy codebase archaeologyInheriting a codebase with no documentation and no original developers left? Claude Code is remarkably good at reading through tangled, undocumented code and producing a coherent explanation of what it does and why β even when the logic is genuinely strange.
02 Large-scale refactoringMigrating from one pattern to another across hundreds of files β say, switching from callbacks to async/await, or from a custom state management system to Redux β is exactly the kind of tedious, error-prone work where Claude Code with Opus model does the heavy lifting.
03 Test suite bootstrappingGreenfield projects rarely have good test coverage early on. Claude Code can read your existing implementation and generate a solid first draft of unit and integration tests β giving your team a foundation to build from rather than a blank file to stare at.
04 PR review assistancePoint Claude Code at a diff and ask it to review for logic errors, security issues, or adherence to your team's style guide. Because it's running on Claude β trained with a strong focus on nuanced reasoning β it catches things that pattern-matching linters miss.
05 Documentation generationNot just README files β Claude Code can generate API documentation from docstrings, create architectural decision records (ADRs) based on your actual code structure, or write changelogs that reflect what genuinely changed between versions.
06 Debugging tricky errorsPaste in a stack trace, give Claude Code access to the relevant files, and ask it to trace the error back to its root cause. It's particularly useful for errors that span multiple abstraction layers β the kind where you know something is wrong but aren't sure which layer is responsible.
Developers evaluating Claude Code often compare it to Cursor, GitHub Copilot, Codeium, and similar tools. The comparison isn't always apples-to-apples, because these tools operate at different levels of abstraction.
The clearest differentiator is the CLI-first, agentic design. Tools like Copilot and Codeium are built to live inside your editor and assist as you type. Claude Code is built to take over and run with a task while you focus on something else β or to tackle the kind of job that's too complex to drive manually through a chat interface.
Is the code I share with Claude Code used to train future models?
Anthropic's API usage policies apply. By default, API data from customers is not used to train models without explicit consent. Review Anthropic's current privacy policy and data usage terms before using Claude Code with proprietary or sensitive source code.
Which Claude model should I use for most tasks?
Claude Sonnet 4.6 is the practical default for most developers. It balances speed, quality, and cost well for everyday coding tasks. Upgrade to Opus for complex architectural work or large-scale refactors where the extra reasoning depth is worth the cost. Use Haiku for high-volume, repetitive subtasks.
Can I use Claude Code with a third-party model through an OpenAI-compatible API endpoint?
No. Claude Code makes direct calls to Anthropic's API and validates that it's connecting to genuine Claude models. You can't proxy a different model through an Anthropic-compatible wrapper β the tool-calling format and response parsing are tightly coupled to Claude's actual output structure.
Can Claude Code commit to git or is it read-only?
Claude Code can run any shell command you authorize, including git operations. It can stage files, write commit messages, and commit β but it will show you what it's about to do and ask for confirmation before taking irreversible actions by default.
If you've used GitHub Copilot or Cursor, you have a rough frame of reference, but Claude Code operates at a different level of abstraction. Rather than completing individual lines or suggesting functions, Claude Code behaves more like a junior developer who lives inside your terminal. You hand it a task in plain English; it reads your files, writes or edits code, runs tests, interprets errors, and iterates β all without you needing to manually copy output between tools.
It's an agentic tool, meaning it takes sequences of actions autonomously. That's a meaningful distinction. Most AI code assistants respond to prompts. Claude Code acts on goals.
This is probably the most common question developers ask when they first encounter Claude Code β and it's a fair one. The short answer is that Claude Code was built from scratch around Claude's specific capabilities, including its extended context window, its reasoning patterns, and its constitutional AI training.
Claude Code relies on Claude's ability to hold massive amounts of context at once β entire codebases, chains of tool calls, multi-step reasoning β without losing track of earlier state. The system prompts and tool-use architecture underpinning Claude Code are tightly coupled to how Claude processes structured inputs and chains thoughts together. Swapping in a different model would break the tool-calling loop that makes the whole thing work.
Claude Code gives you a choice of models depending on your needs β but every option is a Claude model. Here's how the current lineup breaks down:
You can set which model Claude Code uses via the ANTHROPIC_MODEL environment variable or through the --model flag at the command line. For most individual developers, Sonnet is the sweet spot. Opus is worth the extra cost when you're working through something genuinely complex β like migrating a legacy monolith or debugging a subtle concurrency issue across a large codebase.
Rather than a vague list of "AI features," here's a practical breakdown of what Claude Code can handle, based on how it actually works in a real terminal session.
Claude Code doesn't just edit one file at a time. Ask it to rename a function across your entire codebase and it will find every reference, update every call site, and show you a diff before committing anything.
It can run npm test, interpret the output, notice a failing test, trace it back to a specific function, fix it, and re-run the suite β all in one uninterrupted flow.
Point Claude Code at an unfamiliar repo and ask it to explain the architecture. It reads the actual source code β not documentation β and gives you a real mental model of how the system is structured.
Claude Code understands git context. It can review your diff before a commit, write a meaningful commit message based on actual changes, or help you untangle a messy merge conflict with full awareness of both sides.
Give it a module and ask for test coverage. Claude Code writes unit tests that reflect actual edge cases, not just happy paths β because it reads the implementation before it writes the test.
Claude Code supports the Model Context Protocol, letting you connect external data sources β databases, internal APIs, documentation systems β directly into its context. Your company's domain knowledge becomes part of the conversation.
The tool is flexible, but it's not magic. Here are the situations where developers consistently report getting real value out of Claude Code β and where the Anthropic-model architecture actually matters.
01 Legacy codebase archaeologyInheriting a codebase with no documentation and no original developers left? Claude Code is remarkably good at reading through tangled, undocumented code and producing a coherent explanation of what it does and why β even when the logic is genuinely strange.
02 Large-scale refactoringMigrating from one pattern to another across hundreds of files β say, switching from callbacks to async/await, or from a custom state management system to Redux β is exactly the kind of tedious, error-prone work where Claude Code with Opus model does the heavy lifting.
03 Test suite bootstrappingGreenfield projects rarely have good test coverage early on. Claude Code can read your existing implementation and generate a solid first draft of unit and integration tests β giving your team a foundation to build from rather than a blank file to stare at.
04 PR review assistancePoint Claude Code at a diff and ask it to review for logic errors, security issues, or adherence to your team's style guide. Because it's running on Claude β trained with a strong focus on nuanced reasoning β it catches things that pattern-matching linters miss.
05 Documentation generationNot just README files β Claude Code can generate API documentation from docstrings, create architectural decision records (ADRs) based on your actual code structure, or write changelogs that reflect what genuinely changed between versions.
06 Debugging tricky errorsPaste in a stack trace, give Claude Code access to the relevant files, and ask it to trace the error back to its root cause. It's particularly useful for errors that span multiple abstraction layers β the kind where you know something is wrong but aren't sure which layer is responsible.
Developers evaluating Claude Code often compare it to Cursor, GitHub Copilot, Codeium, and similar tools. The comparison isn't always apples-to-apples, because these tools operate at different levels of abstraction.
The clearest differentiator is the CLI-first, agentic design. Tools like Copilot and Codeium are built to live inside your editor and assist as you type. Claude Code is built to take over and run with a task while you focus on something else β or to tackle the kind of job that's too complex to drive manually through a chat interface.
Is the code I share with Claude Code used to train future models?
Anthropic's API usage policies apply. By default, API data from customers is not used to train models without explicit consent. Review Anthropic's current privacy policy and data usage terms before using Claude Code with proprietary or sensitive source code.
Which Claude model should I use for most tasks?
Claude Sonnet 4.6 is the practical default for most developers. It balances speed, quality, and cost well for everyday coding tasks. Upgrade to Opus for complex architectural work or large-scale refactors where the extra reasoning depth is worth the cost. Use Haiku for high-volume, repetitive subtasks.
Can I use Claude Code with a third-party model through an OpenAI-compatible API endpoint?
No. Claude Code makes direct calls to Anthropic's API and validates that it's connecting to genuine Claude models. You can't proxy a different model through an Anthropic-compatible wrapper β the tool-calling format and response parsing are tightly coupled to Claude's actual output structure.
Can Claude Code commit to git or is it read-only?
Claude Code can run any shell command you authorize, including git operations. It can stage files, write commit messages, and commit β but it will show you what it's about to do and ask for confirmation before taking irreversible actions by default.
.png){kind=small}
.png){kind=small}
