Where a coding assistant lives

Four shapes, and what each can see

Coding assistance is the largest paid category in this field, and the products differ far more by where they run than by which model they use. What the tool can see decides what it can do.

Autocomplete in the editor. Suggests the next few lines as you type. Sees the current file and usually a few related ones. Fast, cheap, and it saves real time on boilerplate. It cannot reason about your architecture because it cannot see it.

Chat in the editor. A panel beside your code. Sees whatever you attach plus, in better implementations, an index of the whole project. Good for “explain this”, “write a test for this”, “why is this failing”.

A terminal agent. Runs as a command, reads and writes files across the project, runs your tests, and iterates. This is the shape that changes how the work feels, because it closes the loop: it makes a change, runs the test, reads the failure, tries again. It is also the shape that can do the most damage, since it edits your files.

A hosted agent. You describe a task on a website, it works in a cloud copy of your repository, and it opens a pull request. Good for well-defined, self-contained changes; poor at anything needing judgement about your codebase’s conventions.

Free paths, all usable

You do not need a subscription to have any of this.

• Continue: an open-source extension for VS Code and JetBrains that provides autocomplete and chat against any model, including a local one.
• Aider: an open-source terminal agent that works with git, makes commits per change, and runs against any API or local model.
• Cline and similar open extensions: agentic editing inside VS Code.
• Local models for code. Qwen’s coder models and several others run on a 16 GB laptop and handle everyday completion and small edits well.

The combination of Continue plus a local coding model is a complete, free, offline setup. It is meaningfully worse than a paid frontier assistant on hard problems and entirely adequate for the routine two thirds.

The rule that matters more than the tool

Review every line before it lands. Not because these tools are bad (they are extremely good) but because the failure mode is code that looks right and passes the test you were shown. A subtly wrong function that compiles is more expensive than one that does not.

Two habits protect you, and they cost nothing:

1. Work in small commits. A change you can read in two minutes is a change you can judge. An agent that rewrote forty files in one go is unreviewable, and people accept those because reviewing them is unpleasant.
2. Keep version control between you and the agent. Commit before letting an agent loose. Then git diff is your review tool and git checkout . is your undo. Without that, an agent that goes wrong halfway through is a genuine problem.

Permissions, which is where people get hurt

A terminal agent that can run commands can run any command. Most ask before executing, and most offer a mode that stops asking. That mode is convenient and it is how somebody’s test database gets dropped.

Sensible defaults: let it read freely, ask before writing, always ask before running anything that touches a network or a database. If a tool offers an allow-list of safe commands, spend ten minutes on it once. And be careful with an agent that reads issues, pull request comments or web pages, because text it reads can contain instructions aimed at it, and it has your permissions.

Choosing, briefly

• Working in one file at a time, mostly writing new code: autocomplete covers it.
• Understanding an unfamiliar codebase: editor chat with project indexing.
• Repetitive changes across many files, or test-fix loops: a terminal agent.
• Small, well-specified tasks you want done while you do something else: a hosted agent.
• No budget, or code that cannot leave your machine: Continue or Aider with a local model.

The model matters less than which of these you are in. A middling model in a terminal agent that can run your tests will outperform an excellent model in a chat window that cannot see your project.