What Claude Code is
Claude Code is Anthropic’s coding agent. It runs in your terminal, in a desktop app, in the browser, and as an extension for VS Code and JetBrains editors — and in every case it operates on a real project: reading your actual files, proposing edits, running your tests, and reading the failures.
The difference from a chat window is the loop, not the model. A chat assistant answers a question about code you pasted; an agent reads the file, changes it, runs the thing, sees what broke, and tries again — with each action gated by your approval. For research code that changes what is possible: refactoring a five-year-old analysis pipeline, fixing a LaTeX build that broke on a new machine, or tracing why a figure regenerates differently than it did last year are all tasks that require looking around the project, which is exactly what a chat interface cannot do.
Why researchers use it
- It sees the real project — your directory layout, your half-documented helper scripts, your actual data paths, not a pasted fragment with the context removed.
CLAUDE.md— a plain-text charter in your repository stating what the project is, which folders are sacred, and what the agent must never do. It reads this every session, so your rules persist.- The agentic loop under supervision — run tests, read the failure, fix, repeat. You approve each step, which is both the safety mechanism and the learning mechanism.
- Explaining inherited code — the single most valuable use for most researchers is not writing new code but understanding the pipeline a departed postdoc left behind.
- It works where you work — terminal, IDE, or browser, against the same repository.
Where it fits in a research workflow
Claude Code sits inside your repository, which means everything upstream and downstream is your existing toolchain. Git is the foundation and it is not optional: version control is what makes an agent’s edits reviewable and reversible, and without it you are letting an autonomous process modify files you cannot restore. The habits in Git and GitHub for researchers are the prerequisite, not an optional extra.
From there it composes with the rest of the computational stack: it can run your uv or conda environment, build your Quarto or LaTeX documents, execute a Snakemake pipeline, and commit through GitHub. If you want to understand what an agent actually is before handing one your repository, AI agents explained covers the loop and its failure modes.
Getting started
An afternoon, and the order of the first two steps is not negotiable.
- Put the project under version control first. Commit everything. Git is your undo button, and an agent without one is an agent you cannot roll back.
- Write
CLAUDE.mdbefore running a single task: what the project is for, which directories are read-only, which commands are safe to run, what is forbidden. Ten minutes here prevents the entire class of problems people write threads about. - Start read-only. “Explain what this pipeline does and where it is fragile” teaches you how the agent reasons before you give it permission to change anything.
- The step people skip: read the diff. Every time. The agent proposes, you approve — approval you grant without reading is not supervision, it is a slower way of running unreviewed code.
Claude Code vs the alternatives
| Alternative | Does it better | Pick it if |
|---|---|---|
| Claude | Reasoning about prose, papers, and arguments | Your problem is a manuscript, not a codebase |
| VS Code with an assistant extension | Inline completion while you type, at lower cost | You want suggestions, not an agent that acts |
| GitHub Actions | Deterministic, repeatable automation with no model in the loop | The task is a fixed pipeline, not an open-ended change |
| Doing it yourself | Understanding, which compounds over a career | The code is core to your thesis and you need to know it cold |
That last row is not a joke. An agent that fixes code you do not understand leaves you with a working script and an unanswerable viva question.
Cost, licensing, and your data
Subscription-based, sold through Anthropic’s paid plans with team and enterprise tiers; usage-based API billing is also available. Check current pricing directly rather than trusting a figure in an article.
The data question is sharper here than for a chat assistant, because the agent reads files rather than what you chose to paste. Whatever it looks at goes to Anthropic’s servers under the terms of your plan. For a public analysis repository that is unremarkable. For a repository containing participant data, clinical records, or unpublished results under a data-processing agreement, it is a decision to make deliberately, with your data steward, and probably under an institutional agreement rather than a personal subscription. Read your CLAUDE.md as a governance document, not just a style guide: “never read data/raw/” is a line worth writing.
The honest review
Strengths. For a solo researcher maintaining code nobody else has ever read, Claude Code approximates having a careful engineering collaborator on demand. The compounding effect comes from CLAUDE.md: encode your project’s rules once and every subsequent session starts from them, so the agent gets more useful as the project accumulates conventions rather than less.
Limitations. It amplifies whatever your repository already is. A well-structured project with tests gets faster; a folder of analysis_final_v3_REAL.py files gets confidently rearranged in ways you will struggle to review. There is a genuine learning curve in scoping tasks and setting guardrails, and the failure mode of a badly scoped task is not an error message but a large, plausible, wrong diff. It costs a subscription. And the deepest limitation is not technical: statistics you do not understand remain your problem, and an agent that produces a correct-looking analysis you cannot defend has made your situation worse, not better.
Verdict. Adopt it if your research involves real code, your work is in Git, and you are willing to read diffs. Skip it if your code is a handful of scripts you already understand, if your repository is unversioned, or if the budget would come out of a stipend. The condition that flips the answer is discipline: this tool rewards structure and punishes its absence more sharply than any other in this directory.
CLAUDE.md on day one. And do not use it to produce analysis you cannot explain — an agent accelerates competence, it does not substitute for it, and the viva does not accept “the agent wrote that part”.Common questions
Is Claude Code free?
No. It is available through Anthropic’s paid subscription plans, with team and enterprise options, and can also be used against usage-based API billing. There is no permanent free tier.
Do I need to know how to code to use Claude Code?
You need enough to read a diff and judge whether a change is right. It is a force multiplier on existing competence rather than a replacement for it — used by someone who cannot evaluate the output, it produces code that works until it silently does not.
Does Claude Code work outside the terminal?
Yes. It runs in the terminal, in a desktop app for Mac and Windows, in the browser, and as an extension for VS Code and JetBrains IDEs. All of them operate on the same repository with the same CLAUDE.md rules.