Claude Code: what this agent is and how it differs from a Claude chat

If you ask "what is Claude Code", the short answer is: it is a terminal program in which a Claude model works as an agent over your project. It reads the files it needs, makes edits, runs tests and commands, and looks at the result. Below we cover how this differs from a normal chat and from editor plugins, what it can do, and what working through the API costs.

How an agent differs from a chat

In a chat you work as a go-between: you copy code into a window, get an answer, copy it back, run it, and return with an error. The model sees only what you showed it.

Claude Code works differently. You run the claude command in the project folder and set a task in words, for example "find why the payment test fails and fix it". The agent then acts in a loop:

1. finds the files it needs and reads them;
2. proposes a change and applies it;
3. runs a test or command;
4. reads the output, and if the test still fails, goes back to step 2.

Each such step is a separate request to the model, and it carries the accumulated history. This matters for cost, as described below.

What is Claude Code: the main capabilities

The main capabilities a newcomer should know:

• Working with project files. Reading, searching code, editing several files within one task.
• The terminal. Running tests, linters, git commands. Before potentially dangerous actions the agent usually asks for confirmation; the level of freedom is configurable.
• The CLAUDE.md file. These are project instructions (how to run tests, which code style, what not to touch). The agent reads it at the start of work. See how to write CLAUDE.md↗.
• Extensions. Hooks, subagents and connecting external tools via the MCP protocol. Hooks run your scripts on agent events, subagents take on separate subtasks, and MCP servers give access to external systems.
• Non-interactive mode. The command can be called from a script or CI with a single request and returns the answer in its output.
• Editor integration. Claude Code can run in the built-in terminal of VS Code and other editors.

The exact list of commands and current features depend on the version, so check the official Claude Code documentation.

Claude CLI vs IDE plugins: the difference

"Claude CLI" in search queries usually means the same Claude Code: the command line. Compared with editor plugins:

The choice is a matter of taste. If you live in the terminal or need to run on a remote server, Claude Code is more convenient. If you want to see every change in the editor before applying it, a plugin is closer. A comparison with another popular tool is in the article Claude Code vs Cursor↗.

What Claude Code costs through the API

You can use Claude Code through a subscription or by paying for tokens with an API key. We do not know subscription terms and they change: check the current plan on the service's site. Below we calculate per-token payment through the gateway, where you pay only for what you use.

Suppose a typical "fix a bug" task: 25 steps, an average context of 30 thousand tokens, and a reply of about 500 tokens per step. That gives 750 thousand input tokens (30,000 at each of 25 steps) and 12,500 output tokens (500 at each of 25 steps). Prices as of 18 October 2026, $ per 1M tokens, formula: input × input price + output × output price.

Without cache:

• claude-sonnet-5-5 (0.4 / 2): 0.4 × 0.75 + 2 × 0.0125 = 0.30 + 0.025 = $0.325.
• claude-opus-5-5 (0.8 / 4): 0.8 × 0.75 + 4 × 0.0125 = 0.60 + 0.05 = $0.65.

In the Anthropic format, which is how Claude Code works, the gateway sees the cache and it is billed separately. Suppose that of the 750 thousand tokens, 600 thousand are read from cache, 75 thousand are written to cache and 75 thousand are fresh input:

• claude-sonnet-5-5 (read 0.05, write 0.5): 0.05 × 0.6 + 0.5 × 0.075 + 0.4 × 0.075 + 0.025 = 0.03 + 0.0375 + 0.03 + 0.025 = about $0.12.
• claude-opus-5-5 (read 0.08, write 1): 0.08 × 0.6 + 1 × 0.075 + 0.8 × 0.075 + 0.05 = 0.048 + 0.075 + 0.06 + 0.05 = about $0.23.

The cache shares here are illustrative: in a real session they depend on the task. This is an example calculation, not a promise of specific savings. More on the calculation: agent cost with cache↗. Current prices are GET https://api.zukko.pro/v1/models or the Models page in the dashboard.

How to connect Claude Code with an API key

Create a key in the dashboard↗ and set three environment variables:

export ANTHROPIC_BASE_URL=https://api.zukko.pro
export ANTHROPIC_AUTH_TOKEN=fetch_your_key
export ANTHROPIC_MODEL=claude-sonnet-5-5
claude

The address has no /v1 because Claude Code adds the path itself. The key goes specifically through ANTHROPIC_AUTH_TOKEN. The full instructions with variants for different systems are in the Claude Code documentation↗. If it asks you to sign in at launch, the variables are not set in this terminal window: check them and restart claude.

To change the model for hard tasks, use the same variable: for example claude-opus-5-5, with an intelligence index of 58 against 48 for claude-sonnet-5-5. The index is a general Artificial Analysis score and does not guarantee the result on your project: test both models on your own tasks.

If you want to cap spending, set a daily and monthly dollar limit on the key. This is done on the keys page in the dashboard; once the limit is used up, the gateway answers 429 and the agent stops instead of spending more of the balance.

Who Claude Code suits and who it does not

It suits you if you have a project under git, tasks can be verified by tests or by running the code, and you are ready to read changes before committing. It is convenient to hand the agent routine work that is easy to verify: refactoring, finding the cause of an error, tests, porting code.

It does not suit as a replacement for understanding. If you cannot judge the result, the agent will only create a problem faster. Start with small tasks in a test branch and commit before every big request.

FAQ

Are Claude Code and Claude in the browser the same thing?

The model family is the same, but the environment differs. A chat answers the text you gave it. Claude Code works with your files and commands, which is why it also uses more tokens.

Do I need a separate account for Claude Code?

To connect through the gateway, a fetch_... key and the three variables above are enough. For the terms of official sign-in through a subscription, check the service's site.

Can I use other models in Claude Code?

Through the ANTHROPIC_MODEL variable you can specify any model from the gateway list, since any model can be called in any format. But how well the agent works with a model depends on the model: test on your own tasks.

How much does it cost to try?

Top-ups start from $1. There is $0.50 for subscribing to @zukkopro↗, once per Telegram account: by the calculation above that covers one or two small tasks on claude-sonnet-5-5.