What can Claude Code do?
Claude Code carries out coding work end to end: it finds the relevant files, edits them across a repository, runs your commands and tests, reads what broke and iterates. Around that it handles git — commits, branches, pull requests — reads PDFs and images you point it at, and searches the web when you allow it. What it doesn't do is run without a network, generate images, or replace your editor.
- Multi-file edits
- Yes plans, then changes several
- Runs commands
- Yes tests, lint, build — reads output
- Git workflow
- Yes commits · branches · PRs
- Reads PDFs
- Tested probe returned verbatim
- Sees images
- Tested named a colour it could only see
- Web search
- With permission refuses until approved
- Works offline
- No measured — 3 min, then fails
- Makes images
- No reading ≠ generating
The vendor's list, filtered through what we actually measured.
Anthropic's overview groups the work into building features from a description, debugging from an error message, creating commits and pull requests, connecting external systems through MCP, customising behaviour with instructions, skills and hooks, running agent teams in parallel, piping output through the CLI, and scheduling recurring tasks.
Rather than repeat that, we've tested the parts that people actually ask about as yes/no questions — and each row in the grid above links to the run behind it. Where the answer is no, it's a measured no: the offline test is three timed runs, not an assumption.
The one thing worth stating plainly, because the capability list obscures it: this is an agent, not an assistant that answers questions about code. It changes files on your disk and runs commands on your machine, subject to a permission model you should understand before pointing it at anything you care about.
The work it actually does.
Give it a goal rather than a diff — "write tests for the auth module, run them, fix what fails" — and it locates the files itself, proposes edits across as many as the task needs, executes your test command, reads the output, and iterates until it passes or it's stuck.
That loop is the product. Everything else is a convenience around it.
Three we put probes through, so the answers aren't guesses:
- PDFs — a generated file's probe string came back verbatim in 34 seconds. Scanned documents untested.
- Images — it named the colour of an image whose colour appeared nowhere but the pixels, in 9.7 seconds.
- The web — yes, with a permission grant; without one it refuses rather than answering from memory.
Add MCP and the input surface widens to whatever systems you connect: trackers, docs, internal APIs.
Four honest limits, each with its own page. It doesn't work offline — every turn is a network call. It doesn't generate images. It isn't an editor, so it has no UI for you to write code in. And it isn't free: the install costs nothing, the use does.
None of those are complaints. They're the shape of the tool, and knowing the shape saves you evaluating it for a job it was never going to do.
Different shapes of the same idea.
The capability lists across agentic coding tools converge; where they differ is where the agent lives and what it costs to run.
Frequently asked.
Quick follow-ups people search after this question.