What Can Claude Code Do? A Capabilities Guide

Claude Code is Anthropic's agentic coding tool that runs in your terminal and IDE. It reads and edits files across a codebase, runs shell commands, searches and fetches the web, reads PDFs and images, and connects to external systems through Model Context Protocol (MCP) servers. It is powered by Anthropic's Claude models and integrates with CI pipelines such as GitHub Actions.

What is Claude Code, briefly?

Claude Code is a command-line interface with IDE integrations (VS Code, JetBrains) that lets a Claude model act as an agent inside your development environment. Rather than answering questions in a chat box, it plans work, runs tools, edits multiple files, and verifies results in a loop. For a deeper introduction, see our overview of what Claude Code is for enterprises.

The core idea is the agent loop: the model decides which tool to call, observes the output, and continues until the task is complete. That is what separates it from autocomplete tools.

Reading, searching, and the internet

Can Claude Code read PDFs?

Yes. Claude Code can read PDFs along with images, source files, logs, and structured data. The underlying Claude models support vision and document understanding, so the agent can extract requirements from a specification PDF, parse a diagram, or summarise a design document and then act on it. Large documents are handled within the model's context window, and enterprise plans can use extended context for very large files.

Can Claude Code search the web?

Yes, through tools. Claude Code can perform web search and fetch specific URLs when those tools are enabled, which lets it check current library documentation, resolve an error message against a public issue, or confirm an API signature before writing code. Web access is a tool, not an always-on behaviour, so you control when it is permitted.

Can Claude Code access the internet?

Only when you allow it. Network access happens through explicit tools (web search, web fetch, or MCP servers that call external APIs). In a locked-down enterprise setup, administrators can restrict or disable outbound tools entirely, which matters for regulated environments and air-gapped pipelines.

Which model powers Claude Code?

Which model does Claude Code use?

Claude Code runs on Anthropic's Claude models, drawn from the Opus, Sonnet, and Haiku family. Opus-class models handle the heaviest reasoning and large refactors, Sonnet-class models balance speed and capability for everyday coding, and Haiku-class models suit fast, lightweight tasks. You can select the model per session, and Anthropic ships new versions regularly, so the best practice is to choose by tier rather than memorising a version string.

Can Claude Code use other models or local models?

Claude Code is built for Anthropic's Claude models. It is not a bring-any-model tool, and it is not designed to run arbitrary local open-source models. For enterprises, the same Claude models are available through Amazon Bedrock and Google Vertex AI, so you can run Claude Code against your own cloud account with that provider's security, billing, and data controls. That is the supported path for teams that need a specific deployment or region rather than swapping in a third-party model.

What can Claude Code build?

Can Claude Code build an app?

Yes, within the bounds of an agentic coding tool. It can scaffold a new project, write features across many files, add tests, run them, debug failures, and iterate until the build passes. It is effective at migrations, refactors, bug fixes, and adding endpoints or components to an existing system. It works best with a clear specification and a way to verify its output, such as a test suite or a runnable command. It is not a no-code, push-button app generator; it is an engineering collaborator that produces real code you review and own.

Can Claude Code generate images?

Can Claude Code generate images?

Not natively. Claude Code is a coding agent, not an image generation model. It does not produce illustrations or photos on its own. However, because it is extensible through MCP, you can connect an image-generation service as a tool, and the agent can then call that service as part of a workflow (for example, generating placeholder assets during a build). The generation happens in the connected tool, not in Claude Code itself.

Capability summary

Extending Claude Code with MCP

Where is the Claude Code MCP config?

MCP servers are configured through Claude Code's settings, typically in a project-level or user-level configuration file (commonly a .mcp.json file in the project, plus user and enterprise scopes). You can add servers from the CLI and scope them to a single project or share them across a team. Because configuration is file-based, MCP setups can be committed to a repository so every engineer gets the same database, browser, or internal-API tools. This is how a coding agent reaches your private systems without those integrations being baked into the model.

Will Claude Code replace developers?

Will Claude Code replace developers?

No. It changes how developers work rather than removing the role. Claude Code handles boilerplate, repetitive refactors, test writing, and first-draft implementations, which frees engineers for architecture, judgement, security review, and product decisions. It still needs a human to define the problem, review the output, and own correctness. Teams that adopt it tend to ship faster and spend more time on higher-value work, not fewer engineers doing the same tasks. For a tooling comparison, see Claude Code vs Cursor.

Frequently asked questions

Does Claude Code work in CI/CD pipelines?

Yes. Claude Code integrates with GitHub Actions and similar pipelines, so it can review pull requests, triage issues, and run automated changes on a schedule. See our guide to Claude Code with GitHub Actions.

Can I control which commands Claude Code is allowed to run?

Yes. Permissions are configurable. You can allow safe operations, block destructive ones, require approval for sensitive actions, or run read-only agents that analyse but do not modify code.

Does Claude Code keep my code private?

Code is sent to the model that processes it. When you run Claude Code through Amazon Bedrock or Google Vertex AI in your own cloud account, data handling follows that provider's enterprise controls, which is the usual route for organisations with strict data requirements.

Can Claude Code handle a large existing codebase?

Yes. It is designed for real repositories, using search, file navigation, and context management to work across many files rather than a single snippet.