Claude Code explained, and why it beats an AI plugin in VS Code

A plugin helps you type. Claude Code takes the task, runs it, checks it and reports back.

39%of professional developers using Claude Code at work in May to July 2026, in JetBrains' survey

18%the same figure in January 2026

21%GitHub Copilot in May to July 2026, down from 29% a year earlier

Claude Code is a program that does the task

Anthropic’s documentation calls it “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools.” It runs in a terminal, and also as a desktop app, in a browser and inside an editor.

	An AI plugin, as most people use one	Claude Code
You give it	The line or function you are writing	The outcome you want
It does	“Inline suggestions for completions and suggested next edits,” plus a chat panel	Reads the repository, edits files, runs commands, reads the errors and fixes them
It stops	After each suggestion	When the task is done, or when a check you named passes
It runs	In an editor window you have open	Anywhere a terminal runs, with you or without you

1. You ask"Build a firm-quarter panel of total assets and revenue for the firms in sample.csv from SEC filings. Report how many firms have no data."

2. It worksIt writes the download script, runs it, hits a rate limit, adds a delay, reruns, builds the panel and runs the tests.

3. You reviewYou read the row counts, the test output and the commits it made.

Terminal~/sec-panel

$ claude
 
> Build a firm-quarter panel of total assets and revenue for the firms
  in sample.csv from SEC filings. Report how many firms have no data.
 
● Read(sample.csv)
● Write(pull.py)
● Bash(python pull.py)
  ⎿ HTTP 429: too many requests
● The SEC allows 10 requests a second. I'll add a delay and a cache.
● Edit(pull.py)
● Bash(python pull.py)
  ⎿ wrote data/panel.parquet
● Bash(pytest -q)
  ⎿ 5 passed

⏵⏵ auto mode on (shift+tab to cycle)

An illustration of a session, not a recording. You type one request. The lines marked with a dot are the agent's own actions.

Developers moved to the labs’ own agents in six months

JetBrains surveyed more than 15,000 professional developers for its Developer Ecosystem Survey 2026.

Tool	Earlier	May to July 2026
Claude Code	18% in January 2026	39%
Codex	3% in January 2026	16%
GitHub Copilot	29% a year earlier	21%
Cursor	18% in January 2026	12%

The two tools that grew are the agents built by Anthropic and OpenAI, the companies that train the models. JetBrains’ own summary is that “GitHub Copilot has lost its leadership.”

The gain is in the work that runs while you are away

You want	Claude Code gives you	Since
No permission prompt before every command	Auto mode, where a classifier reviews each action	March 2026, and the default from 14 August 2026
Work that continues until a check passes	`/goal`, followed by the condition	May 2026
A session that outlives your working day	Start it inside `tmux` on a machine that stays on, type the goal and detach
The same job every morning	`/loop 1d` followed by the task, in a session you leave open
An agent that knows your project	An instruction file and skills, read at the start of each session

The SEC filings guide puts the first three together: one typed goal builds a panel, tests it and commits it. Instruction files and skills covers the last row.

Switch to Claude Code fully

You want to be at the frontier. It ships in Claude Code every week, and a researcher who waits for an editor plugin to catch up is choosing to be behind.

Do not keep one foot in the editor. Anthropic calls the terminal program “the full-featured CLI,” and new abilities arrive there week after week. Its documentation has published a “What’s new” digest for 24 separate weeks of 2026.

Week of	Shipped in Claude Code
23 March 2026	Auto mode
30 March 2026	Computer use in the terminal
13 April 2026	Opus 4.7 and the `xhigh` effort level
11 May 2026	`/goal`
3 August 2026	Auto mode becomes the default

It remembers youAuto memory "lets Claude accumulate knowledge across sessions without you writing anything." Correct it once and the note is there next time.

You write plain EnglishThere is no syntax to learn. You state the outcome and the check that proves it.

It goes where the data isA terminal program runs on the cluster you reach over SSH. A plugin runs where the editor is installed.

VS Code’s documentation, dated 7 October 2026, now describes agents of its own and lists Claude among the agents it can host. That is the editor following the agent. Go to the source: use Claude Code itself, on Anthropic’s best model, and take every update the week it ships. Agent-first coding makes the case for paying for the top of the range.

A stock agent already writes a paper that reads like an average submission

In a May 2026 study, Claude Code with a light scaffold ran the whole research loop from idea to manuscript. On manuscript-only review its papers scored highest of three agents and matched the weighted-average human submission to ICLR 2025. Once reviewers opened the code, scores fell and no paper cleared a top venue’s bar (the brief).

A plugin user has not seen an agent attempt that at all.

Install it and hand it one task

curl -fsSL https://claude.ai/install.sh | bash
cd your-project
claude

That is the command for macOS, Linux and WSL. The documentation gives the Windows versions. Then type the outcome you want and a check that proves it, and read what comes back.