Everything you need to get the most out of Cinnacode — from your first install to advanced multi-file editing.
You can create a Cinnacode account either from this website or directly inside the extension — both lead to the same place.
Use Get started on this site (guest signup, 200 free credits, no card needed), or install the extension first and it'll prompt you to register on first launch.
Search "Cinnacode" in the VS Code Marketplace, or use the download link on your dashboard.
Open the Command Palette (Cmd+Shift+P), run AI: Login with GitHub, and paste the access key from your dashboard when prompted.
Open the Cinnacode icon in the Activity Bar, type a question in the sidebar, and press Enter.
For multi-step tasks — "add error handling to every route in this file," "write tests for this module" — Agent Mode plans the work, then writes files and runs commands to get it done, one approved step at a time.
1. Open the Command Palette (Cmd+Shift+P) and run AI: Start Agent Task.
2. Describe what you want done, in plain English.
3. Cinnacode proposes a plan and the first step — review it.
4. Approve each file write or command individually; nothing runs without your sign-off.
5. The task continues step by step until it's done, or you click Stop Task.
Agent Mode can write files and run commands on your machine. Approving an action means you've reviewed it — not that Cinnacode has verified it's safe or correct. Read each proposed step before approving it, the same way you'd review a diff before merging it.
Choose a model tier (Fast / Balanced / Capable) right in the Agent Mode panel — the same tradeoff as sidebar chat's model selection, scoped to this task.
Run a small model on your own machine — via Ollama, free and open-source — to answer simple questions without spending credits or needing an internet connection. Available in both Sidebar Chat and Agent Mode's reply box. Cinnacode always checks first whether a question genuinely needs Claude's real reasoning; anything requiring real analysis, code generation, or multi-step logic still goes to Claude automatically, even with Local selected.
1. Install Ollama (macOS, Windows, or Linux).
2. Pull the model: ollama pull qwen2.5:1.5b-instruct-q4_K_M (a one-time, ~1GB download).
3. Open the Command Palette (Cmd+Shift+P) and run AI: Enable Local Mode.
4. Switch the provider dropdown from Claude to ⚡ Local, in either Sidebar Chat or Agent Mode's reply box.
Local Mode answers come from a small model on your own machine, not Claude — apply the same review scrutiny as you would to any AI-generated output.
Type any of these at the start of a chat message to get a focused response, using whatever file you currently have open (or your current selection, if you have one) as context.
| Command | What it does |
|---|---|
| /fix | Returns the complete corrected version of your file, plus a short list of what was fixed. |
| /explain | Explains what the code does, in plain terms. |
| /improve | Suggests improvements for readability, performance, and best practices. |
| /test | Writes unit tests covering the main functionality and edge cases. Automatically detects your project's test framework (Jest, Pytest, JUnit, etc.). |
| /edge-cases | Goes deeper than /test — specifically hunts for boundary values, empty/null inputs, and adversarial scenarios most people forget to test. |
| /coverage | Compares your source code against its existing test file and tells you specifically what's not covered yet. |
| /docs | Writes documentation or comments for the code. |
| /find-usage | Real, mechanical search — finds every exact place a function or variable is used across your whole project, not a guess. |
| /review | Reviews your current uncommitted git changes before you commit — checks for bugs, consistency, and leftover debug code. |
| /readme | Generates a README.md for your project based on its actual structure. |
| /changelog | Generates a changelog entry from your recent commit history. |
You can add extra instructions after the command, e.g. /fix also handle empty arrays.
When Claude mentions a real file in your project by name — even one you never opened or @ mentioned — a small button appears letting you pull in its actual content with one click, instead of typing @filename yourself. If several files are mentioned at once (common for broader "how does X work" questions), you can attach all of them together in a single follow-up.
The fastest way to change a specific piece of code without leaving your cursor position or opening the sidebar.
1. Select the code you want to change.
2. Press Cmd+K (Mac) or Ctrl+K (Windows/Linux).
3. Type what you want changed, in plain English.
4. A diff opens automatically showing the proposed change.
5. Confirm to apply it to your file, or dismiss to cancel.
Cmd+K only activates when you have text selected — it won't interfere with the shortcut elsewhere in the editor.
Every code suggestion — whether from chat or Cmd+K — goes through the same safety pattern: you always see the diff before anything touches your file.
Opens VS Code's native diff view comparing your current code against the suggestion.
Stays disabled until you've viewed the diff at least once. Clicking it shows one more confirmation dialog before writing the change to your file — nothing is ever applied silently.
Pull other files into a chat conversation, not just the one you have open.
Type @ followed by a few letters of a filename in the chat box — this searches your entire workspace live, not just files you've already opened, and shows matching results in a dropdown. Click one to attach it, then finish your question.
Mentioned files are read fresh each time you send a message, capped at a reasonable size per file to keep responses fast and affordable.
Attach a document — a requirements doc, a spec, anything text-based — and ask Cinnacode to work from it.
Click the + button next to the chat input, choose a file, and it attaches to your next message. Currently supports .txt, .md, .json, and .csv — PDF and Word support are planned next.
Cinnacode can compare a document's requirements against your actual code — pointing out what's already implemented, what's still missing, and proposing the code for what's left. It won't create or write files on its own; you review and apply each suggestion the same way as any other response.
Choose the right balance of speed and capability for what you're working on, right from the chat panel.
| Tier | Best for |
|---|---|
| ⚡ Fast | Quick questions, simple edits — fastest responses |
| ⚖️ Balanced | Default — the best mix of speed and quality for everyday coding |
| 🧠 Most Capable | Complex logic, architecture decisions, tricky bugs |
Your selection is remembered across sessions. Inline completions always use the fastest tier automatically, regardless of this setting, since they don't need deep reasoning.
Cinnacode uses prepaid credits instead of a subscription — buy once, spend at your own pace.
| Action | Cost |
|---|---|
| Chat message | 28 credits |
| Inline completion | 6 credits |
| Signup bonus (guest or GitHub) | 200 free credits |
Credits are valid for 9 months from purchase. See the full Terms and Refund Policy for details.
// video walkthroughs coming soon
You'll see a message in the chat panel with a link to buy more. If you're on a guest account, you'll also get a prompt to log in with GitHub for 100 more credits.
Code you're actively working with is sent to Anthropic's Claude API to generate a response — that's how the AI assistance works. We don't use your code to train any model, and we don't sell it. See our Privacy Policy for full details.
You can have both, but they're currently separate accounts with separate credit balances rather than one merged account. Logging in with GitHub from a guest session creates a new account rather than upgrading the existing one.
If you registered with an email and password (not GitHub), go to login.html and click "Forgot password?" — you'll get a reset link by email, valid for 30 minutes. GitHub-linked accounts don't need this; just log in with GitHub again.
Credits that were successfully delivered aren't refundable for change-of-mind, but genuine billing errors (double charges, failed delivery) are covered. See the full Refund Policy.
Run AI: Send Feedback from the Command Palette inside the extension, or reach out via the Contact page.