CodeSprite

AI ghost text for VS Code. Inline completions, AI commands, and commit messages — bring your own key, open-source.

⚠️ CodeSprite is an early prototype. Have an idea, found a bug, or want to contribute? Head over to GitHub and open an issue or PR. Feedback welcome.

CodeSprite is a lightweight VS Code extension that adds three AI coding features without locking you into a specific provider. Choose your preferred AI provider from the built-in presets or bring any OpenAI-compatible endpoint.

• Inline completions — ghost-text suggestions appear as you type, in any language.
• AI Command modal — trigger a natural-language instruction (Ctrl+Shift+I) and insert or replace code at the cursor.
• Commit message generator — one click in the Source Control panel turns your staged diff into a Conventional Commit message.

Built with TypeScript + esbuild. No runtime dependencies. MIT licensed.

Features

1. Inline completions

Just type. After the configured debounce delay, CodeSprite requests a completion and renders it as ghost text. Press Tab to accept. Suggestions are de-duplicated against the code that already follows your cursor, so you never get duplicate lines.

If the status bar shows $(warning) CS, no API key is set. If it shows $(error) CS, the last request failed — hover for details.

2. AI Command modal

Press Ctrl+Shift+I (macOS: Cmd+Shift+I) with a text editor focused. Type a natural-language instruction, for example:

• create a new variable called abc

• convert this loop to a list comprehension

• add a docstring

• No selection — the generated code is inserted at the cursor.

• With a selection — the selected text is replaced with the generated code.

Inserted code is briefly highlighted so you can review it; the highlight clears on the next edit, cursor move, or after 5 seconds. Use Ctrl+Z to undo if it's not what you wanted.

3. Commit message generator

Open the Source Control panel (Ctrl+Shift+G). Click the sparkle/lightbulb icon in the SCM title bar. CodeSprite reads your staged changes (git diff --cached), falls back to unstaged changes if nothing is staged, and writes a Conventional Commit message into the SCM input box.

The commit message prompt extracts a user-story number from the branch name when it follows the 1234-description pattern, prefixing the message with [1234].

Quick start

Requires VS Code ^1.80.0 and an API key for your chosen AI provider.

1. Install — search for CodeSprite in the Extensions view (Ctrl+Shift+X).
2. Configure — open Settings (Ctrl+,), search for codesprite, and set at least:

   {
      "codesprite.provider": "openai",   // or: anthropic, gemini, mistral, xai, custom
      "codesprite.apiKey": "sk-your-key-here",
      // apiBaseUrl and model are optional — the 'openai' preset defaults to
      // https://opencode.ai/zen/v1 with minimax-m2.7. Other presets use their own defaults.
    }
   

3. Use it — just type for inline completions, press Ctrl+Shift+I for the AI Command modal, or click the sparkle icon in the Source Control panel to generate a commit message.

Click the CS indicator in the status bar to toggle all features on/off at once.

Status bar

A CS indicator in the bottom-left status bar reflects the current state:

Configuration

Open Settings (Ctrl+,) and search for codesprite, or edit your settings.json directly.

Providers

CodeSprite ships with built-in presets for the major hosted AI providers. Set codesprite.provider to pick one. The extension automatically uses the correct API endpoint, authentication method, request format, and context window for each provider.

Leave apiBaseUrl and model empty to use the provider defaults. Override them to point at a different endpoint or model.

* The openai preset defaults to https://opencode.ai/zen/v1 (the opencode.ai BYOK service) with model minimax-m2.7. To use literal OpenAI, set apiBaseUrl to https://api.openai.com/v1 and model to gpt-4o or gpt-4o-mini.

Custom provider

Use "codesprite.provider": "custom" for any OpenAI-compatible endpoint — OpenRouter, Together, Groq, Ollama, LM Studio, OpenCode, or self-hosted. The custom preset uses the standard OpenAI chat completions format (/chat/completions, Bearer auth, SSE streaming). Just set apiBaseUrl and model to match your endpoint.

Example: Gemini setup

{
  "codesprite.provider": "gemini",
  "codesprite.apiKey": "AIza..."
  // apiBaseUrl and model are optional — defaults are used
}

Note: Gemini sends the API key as a URL query parameter (?key=...), not as an HTTP header. If you're behind a proxy that logs query strings, consider using a VPN or a different provider.

Example: Anthropic setup

{
  "codesprite.provider": "anthropic",
  "codesprite.apiKey": "sk-ant-..."
  // system prompt is sent as a top-level field, not in the messages array
}

Example: Ollama (local)

{
  "codesprite.provider": "custom",
  "codesprite.apiBaseUrl": "http://localhost:11434/v1",
  "codesprite.apiKey": "ollama",  // Ollama ignores the key but requires a non-empty value
  "codesprite.model": "codellama:7b"
}

Installation

From a .vsix package

code --install-extension codesprite-0.1.0.vsix

From source

git clone https://github.com/mmilidoni/vscode-codesprite.git
cd vscode-codesprite
npm install
npm run package      # produces a .vsix in the project root
code --install-extension *.vsix

Build from source

npm install
npx tsc --noEmit        # type-check (esbuild does NOT run tsc)
npm run build           # bundle to dist/extension.js with sourcemaps
npm run watch           # rebuild on change
npm run vscode:prepublish   # production minified bundle
npm run package         # production bundle + .vsix

Type-check first, build second. A clean build does not guarantee type safety — esbuild strips types without checking them. Always run npx tsc --noEmit before committing.

Press F5 in VS Code to launch the extension host (runs the default build task, then opens a new window with CodeSprite loaded).

Architecture

Single entrypoint: src/extension.ts → activate() wires up the three features. Each feature owns its own abort controller and shares a single status bar item.

All HTTP traffic flows through one streaming pipeline that auto-detects SSE vs. JSON responses, supports early stream termination, and handles servers that omit Content-Type. See AGENTS.md for the full contributor guide and gotchas.

Privacy

CodeSprite sends the code around your cursor (prefix + suffix, trimmed to your configured token budget) and your natural-language instructions to the API endpoint you configure. No telemetry is collected by the extension itself. Review your provider's data policy before use.

License

MIT © mmilidoni