A web chat can only reason about what fits in the message box. Claude Code reasons about your repository. Everything else follows from that: how large the files can be, whether the code runs, and what you pay for the model to see it.
The upload wall is a design decision
Chat apps cap attachments for sensible reasons: storage, abuse, and a UI for everyone. In practice the caps bite. Long PDFs get truncated, images get downscaled, and a repository has to be zipped into an archive the model cannot navigate. You summarize your code before the model sees it.
What a local CLI changes
- File access by path. It opens the files it needs, greps for a symbol, and follows an import across directories.
- Real commands. It runs the tests, linter, and build, then reads the error instead of guessing from your description.
- Edits in place. Diffs land in your working tree, reviewed with git, not in a transcript you copy back.
- Nothing leaves disk. You send context, not a duplicate of the repository.
The practical effect: the model can gather evidence instead of trusting your summary of the failing test.
Artifacts are demos; files are deliverables
Artifacts are a good way to show a chart, a prototype, or a single-page tool. They also run in a sandbox with no access to your dependencies, environment variables, database, or private packages. A component that renders can still fail on a missing import or an undefined secret. Claude Code runs where the code will actually run.
The cost difference cuts both ways
Chat feels free because it is priced by subscription, but that subscription has usage limits and a ceiling on context. Claude Code can be billed by subscription or by API tokens, and an agentic session is a loop: every turn resends the conversation, the file reads, and the command output. A 2,000-line stack trace is input you pay for again next turn.
- Turns — each one resends the transcript. Long sessions cost more than several short ones.
- Tool output — test logs and file reads add up quickly.
- Model tier — routine edits do not need the largest model.
- Cache hits — stable prefixes can be billed at a lower rate.
The counterweight is that one session often replaces an hour of copy-paste and finishes tasks a chat cannot attempt. Price the task, not the message. Forty turns can beat three days of manual work.
Task: fix the failing checkout test. Scope: src/checkout/**, tests/checkout/**. Do not touch: pricing engine, migrations, public API. Steps: reproduce with npm test -- checkout, make the smallest fix, rerun the suite. Report: files changed, root cause in two sentences, test summary. No refactor.
Scope and "do not touch" lines beat a paragraph about your architecture, and keep the diff reviewable.
When chat is still the better tool
- 1One-off questions. A regex, a SQL join, a concept explained.
- 2Code you can paste. A 30-line snippet beats pointing a CLI at it.
- 3Work outside the repo. Screenshots, documents, slides, and drafts.
- 4Restricted machines. No shell access, credentials, or install.
- 5Thinking before committing. Brainstorming without touching files.
A decision rule that takes ten seconds
Ask what evidence the answer needs. If the model must read more than three files, run a command, or see the real error, use Claude Code. If the answer lives in what you can paste, use chat. If unsure, start in chat and switch the moment you catch yourself describing code instead of showing it.
The tool sets the ceiling on evidence. A chat sees your summary. A CLI sees the stack trace.
Frequently asked questions
Can I use Claude Code without a terminal?
Claude Code runs from the terminal, and editor integrations offer the same local file access. It still works against the project on your machine, not an uploaded copy.
Is Claude Code cheaper than a chat subscription?
Not per message. It can be cheaper per finished task because it removes the copy-paste loop and the retries caused by missing context. Compare the cost of a completed task.
Do artifacts count as running code?
They run in a sandbox with limited access to your project. Useful for demos and visual checks, but not a substitute for a build and test run in your own environment.
What can Claude Code read?
Files in the project directory you start it in, plus anything you grant access to. It asks before running commands or editing files, depending on your permission settings.
All content is original. The comparison was informed by the CLI versus chat module of a Claude Code course; no course text or video is reproduced. Cost figures are computed at build time from the TokenCalc AI catalog and will move when prices change.