An AI coding agent doesn’t have to be a separate chat window you copy-paste between. Continue is an open-source extension that drops the model directly into VS Code or JetBrains, with your project files in context. This run-book gets it running in about 20 minutes and flags the config traps that eat a whole afternoon.
If you’d rather start with a hosted option, the GitHub Copilot documentation covers the managed path; this run-book focuses on Continue because it runs against local models too and keeps your code on your machine.
What you need before you start
- VS Code or a JetBrains IDE (IntelliJ, PyCharm, GoLand).
- An internet connection for the initial install. For a local model you’ll also want 8 GB+ free RAM.
- Either an API key for a cloud model, or Ollama installed for a local one.
The Continue quickstart is the canonical setup reference this run-book condenses.
Step 1 — Install the extension
In VS Code, open the Extensions view (Cmd+Shift+X), search “Continue”, and install it. In JetBrains, use the Marketplace tab in Settings → Plugins. After install, a Continue sidebar appears on the right. That’s your agent surface.
Step 2 — Choose a model backend
Continue is backend-agnostic. Two sane starting points:
Cloud (fastest first run): open the Continue config and set a provider block with your API key. Continue reads ANTHROPIC_API_KEY / OPENAI_API_KEY from the environment, so export it in the terminal you launched the editor from.
Local (private): install Ollama, pull a coder model, and point Continue’s provider at localhost:11434. Nothing leaves your machine.
Don’t try to tune both at once. Pick one, prove it works, then switch.
Step 3 — Run your first real task
Open a messy function in your project. Select it, then in the Continue sidebar type: “Add type hints and extract the inner loop into a named helper. Keep behavior identical.”
Watch what happens: Continue reads the selection, pulls in related files it thinks are relevant, proposes an edit, and shows a diff you can accept or reject per-hunk. Your job is to review the diff like a senior teammate wrote it — because the model will occasionally “fix” a bug that wasn’t there.
Step 4 — Confirm it’s grounded in your code
The test of a coding agent is whether it cites your symbols, not generic ones. Ask it: “Where is retry_with_backoff called, and what breaks if I change its signature?” A grounded agent names real files and call sites. If it hallucinates function names, your context window is too small or the model is too weak — bump one, not both.
Config mistakes that waste an afternoon
- Editor launched from a GUI, key exported in a shell. Continue can’t see
ANTHROPIC_API_KEYif the editor process didn’t inherit it. Launch the editor from the terminal, or set the key in the Continue config file directly. - Context window smaller than the file. A 4k window chops your module in half and the agent edits blind. Match the window to the model.
- Accepting diffs without reading them. The single biggest source of “the agent broke my code” is unread accepts. Review every hunk.
- Local model too small for code. A 3B model writes plausible-looking but non-running code. Use at least a 7B coder quant for real work.
Where to go next
Once edits flow, add a .continueignore so the agent skips node_modules and build output, and configure autocomplete to trigger on a keybinding instead of on every keystroke. At that point Continue stops being a toy and becomes part of your normal workflow.
