AI coding: what it is, how it works, and the tools in 2026

AI coding is writing software with a language model doing part of the work: finishing the line you started, answering a question about your code, or taking a whole task, reading the repository, editing files, running the tests and coming back when it’s done. The three are different products with different risks, and most “which AI coding tool” arguments are people comparing one kind to another. This page sorts them, says what each costs, and shows what the agent kind looks like when it runs.

The three kinds of AI coding

Autocomplete changes how fast you type. Agents change what you do all day: you write the task and review the result, and the agent does the typing. App builders skip the code entirely, until you need to change something they didn’t anticipate.

Most products now span two rows. Cursor is an editor with autocomplete and an agent. GitHub Copilot is autocomplete, a chat, a terminal agent and a cloud agent that opens pull requests. Cline and Kilo Code started as editor agents and now ship a CLI, a desktop app and cloud runs. The row that matters is the one you spend your day in.

What AI coding looks like with an agent

A real task, as it runs in a terminal agent:

> Add a --json flag to the export command and cover it with a test.

  Read   src/cli/export.ts
  Read   src/cli/export.test.ts
  Edit   src/cli/export.ts            (+14 −2)
  Edit   src/cli/export.test.ts       (+21)
  Run    npm test -- export           ✓ 9 passed
  
  Added --json to export; output goes through JSON.stringify with
  2-space indent. One new test checks the flag and the shape.

Each line is one turn of a loop: the agent asks the model what to do, the model answers with a tool call, the agent runs it and feeds the result back. The loop stops when the model answers with text instead of a tool call. Every edit and command went through the agent’s permission rules first; here, reads and the test run were allowed without asking, and the edits were approved once for the session. What is an AI coding agent? shows one turn of that loop in the JSON the model actually sees.

Which AI coding tool for which job

The ranking behind the agent rows, with the benchmark data, is in best AI coding agents. The licensing, sign-in and sandbox matrix for every agent is in the AI coding harness comparison.

Is AI coding free?

The tools are often free. The model usually isn’t.

• Free plans exist for the big ones. GitHub Copilot, Cursor and ChatGPT (which includes Codex) each have a free tier with limited usage. OpenCode, Cline, Kilo Code, goose, Aider, Pi and Contenox are free and open source; you bring the model.
• A local model costs nothing per token. OpenCode or Codex with a model from Ollama runs with no subscription and no bill, at the speed of your GPU. Best Ollama models for coding lists what fits in 8, 16, 24 and 48 GB.
• The bill is the model, not the tool. An agent resends the whole conversation on every step. One eight-prompt session of ours sent about 300 million input tokens: about $80 on Claude Fable 5.1 with every optimization on, under $1 on DeepSeek V4 Flash (since replaced by V4.1 Flash). Same task, same tool. AI coding costs shows where the tokens go and how to cut them.

Plans and limits change monthly; the one page that keeps them side by side is AI coding subscriptions compared.

How difficult is AI coding?

Typing is no longer the skill. Three others are:

1. Writing the task. “Add a —json flag and cover it with a test” gets a good result. “Make the export better” gets a guess. The agent does the smallest thing that satisfies the words; if you want error handling and an empty state, say so. How to write good tasks for coding agents.
2. Reading the result. You review a diff instead of writing it. Reviewing well is harder than it sounds when the diff is clean, confident and wrong in one line.
3. Setting the rules. Which commands may run without asking, which files may change, what’s blocked. Set once per project; otherwise you either approve every step until you stop reading them, or switch approvals off. Why does my coding agent keep asking for permission? has the settings for each agent.

The repository matters as much as the prompt. A build command that works, tests that run in seconds, and an AGENTS.md with the three things the agent keeps getting wrong raise every tool’s results. That work has a name, harness engineering, and it’s where the teams getting the most out of AI coding spend their time.

Where AI coding goes wrong

• It declares victory too early. The first passing test ends the loop. Make the agent run the whole suite, or hand the result to a second checker.
• It runs something you didn’t want. git push --force, rm -rf, a migration against the wrong database. Each agent has a skip-all mode, and each agent’s docs tell you not to use it outside a container. What those modes do, agent by agent, is in YOLO mode.
• It reads what it shouldn’t. Your .env, your SSH keys, your cloud credentials are in the working directory too. Keep secrets away from AI agents.
• It costs more than you expected. See the token numbers above; long sessions and big files are the usual cause.

The fix for the middle two is the same: permission rules that say yes to the routine and no to the dangerous, kept on. A sandbox limits the damage when the rule is wrong. How to sandbox AI coding agents covers what each agent offers on macOS, Linux and Windows.

Is AI coding allowed?

Three separate questions hide in that one.

• May you use the code it writes? Every major tool’s terms let you use the output; read your provider’s terms for the exact grant. Whether purely generated code is itself protected by copyright is unsettled in most jurisdictions, which matters if you plan to assert copyright on it.
• Where does your code go? To the model provider, on every request, unless the model runs locally. Business plans from Anthropic, OpenAI, GitHub and Google state that they don’t train on your data; consumer plans differ. Companies with code that can’t leave the building run local models, or an on-premises product such as IBM Bob.
• Does your employer allow it? Many don’t, or allow one tool. The agent’s permission log is what an audit will ask for.

How do I start?

Pick the agent that matches the plan you already pay for, and install it:

npm install -g @anthropic-ai/claude-code

npm install -g @openai/codex

npm install -g opencode-ai

Then, in your project, set the permission rules before the first task. Each agent’s install, first run and permission setup is in terminal coding agents. On Windows, read AI coding agents on Windows first; sandboxes differ there more than anywhere else.

Contenox

Contenox is the agents row for people who want the rules visible: a desktop app where you open a project folder, pick a provider (a ChatGPT subscription, an API key, a local model, or Claude Code hosted inside the app), and a policy engine decides what the agent may run. Builds and tests run without asking; edits and risky commands ask; secrets are blocked. Download it here.

Related guides

• What is an AI coding agent?: one turn of the loop, in JSON
• Best AI coding agents: the ranking and the criteria behind it
• AI coding costs: where 300 million tokens come from
• What is an AI harness?: the software that turns a model into an agent