Orbit AI 🪐

Always watching your code.

Orbit AI is a powerful VS Code extension that provides agentic AI chat features offline, online, and via custom API keys. Whether you are using built-in local models for absolute privacy, or leveraging cloud APIs for maximum speed, Orbit delivers intelligent code explanations, semantic search, AI-driven code reviews, and interactive chat.

🚀 Features

🤖 AI Chat & Smart Apply

Interact with your codebase using a chat interface. Ask questions, request code snippets, or discuss architectural decisions.

• Agent-style chat: Replies stream in live. Each action the agent takes (reading a file, searching, running a command, editing) shows as a step you can expand to see its input and result. Press Esc or the Stop button to cancel.
• Inline change review: Edits appear directly in your file, with removed lines in red and added lines in green. Accept or reject each change with the buttons above it, or everything at once with Ctrl + Enter / Ctrl + Backspace. The file is saved only after you decide.
• Safe "add" requests: When you ask Orbit to add code, it never deletes code you already have, and it tells you honestly when a change could not be made.
• Auto-Apply Detection: If the AI answers with a code block for your active file instead of editing it, Orbit still offers the inline review.
• Attachments: Attach any file (including files outside the workspace) or an image to a message. Images are sent to cloud and local-server models; the built-in offline model cannot read images and says so.
• Context-Aware: Orbit understands your active file and deep dependencies.
• Semantic Search: Search your entire workspace with semantic understanding.

🐚 Terminal Integration

• Run Commands: AI-suggested bash/shell commands get a "Run ▶" button. Orbit executes them with your confirmation and captures the output.

📝 AI Code Review

Get instant feedback on your code or Git changes.

• Review Code: Analyze the current file for bugs, improved practices, and readability.
• Review Git Changes: Automatically review uncommitted changes in your git repository.
• PR Integration: Connect AI review directly to Git branches and Pull Requests.

✈️ Orbit Pilot (Agentic Capabilities)

• Background Analysis: Orbit silently analyzes your active files and reports issues to the VS Code Problems tab.
• Auto-Fix Engine: Automatically repair low-risk issues as you type.

🧠 Decision Journal

• Memory: AI-driven edits and auto-fixes are logged to .orbit/decisions.db, using the SQLite engine built into VS Code's Node.js runtime (no native module to compile).
• Team Sync: Share the journal with your team. The file method writes .orbit/journal.jsonl; the git method keeps the journal on a separate orbit-journal branch and never touches your own branches, index or uncommitted work.

📋 Prerequisites

Orbit needs VS Code 1.101 or newer. It works out of the box in the default cloud mode. Other modes need:

• offline: Node.js (for npm). On first use Orbit asks, then installs node-llama-cpp into ~/.orbit/runtime (several hundred MB) and downloads the default Qwen2.5-Coder-7B model to ~/.orbit/models (several GB).
• Local semantic search: the first index also asks to install @huggingface/transformers (about 100 MB) into ~/.orbit/runtime.
• local-server: a running OpenAI-compatible server such as LM Studio (see orbit.localServerEndpoint).
• custom, groq, openrouter: an API key for the provider.

⚙️ Configuration

Choosing a provider: click the provider button in the chat header, the Change link on the welcome screen, or the Orbit item in the status bar. Pick a provider, paste an API key if it needs one, and click Test connection. Keys entered there are stored encrypted in VS Code's Secret Storage. You can also set everything below in Settings.

⌨️ Keyboard Shortcuts

Shortcuts apply when the editor has focus (Cmd replaces Ctrl on macOS).

🧭 Commands

Open the Command Palette (Ctrl + Shift + P) and type Orbit:

🔒 Security

Orbit lets an AI model act on your workspace, so file and command access is restricted:

• Workspace-restricted tools. The agent's ls, read and apply tools resolve every path against the open workspace folder. Paths that escape it (.., absolute paths elsewhere, or symlinks that resolve outside) are rejected with Access denied.
• Inline apply validation. Proposed edits are checked against the workspace before anything happens. They are shown inline in the file and saved only after you accept them. Rejecting restores the original text, and a new file that is rejected is deleted again. If you type in the file during the review, your edits win and the proposal is dropped.
• Terminal command hardening. Commands never run silently: each one shows a modal prompt with the command and working directory and runs only on Run. The working directory must be inside the workspace, empty commands are refused, and captured runs have a 30-second timeout and a 1 MB output cap.
• Tool argument validation. Malformed tool arguments (null, arrays, empty paths, non-string commands or queries, empty apply code) are rejected before reaching a tool.
• Escaped webview output. Structured AI responses are HTML-escaped before being rendered in the chat.
• PR comments. Posting a review to a pull request goes through the GitHub CLI (gh pr comment, using your existing gh login) with arguments passed directly, not through a shell. It is controlled by orbit.prIntegration.enabled.
• Prompt size cap. Prompts over 200,000 characters are truncated before they are sent to a model.
• Journal sync isolation. Git sync writes only to the orbit-journal branch using git plumbing commands, runs git without a shell, skips malformed journal lines, and uses a lock so two windows cannot sync at once.

In cloud, custom, groq and openrouter modes your prompts and code context are sent to the chosen provider. Use offline or local-server mode to keep everything on your machine.

💻 Development

1. Clone the repository.
2. Install dependencies: npm install
3. Build UI: npm run build:ui
4. Compile Extension: npm run compile
5. Run: Press F5 to open the Extension Development Host.

Testing

• Unit tests: npm test
• Webview tests: npm test --prefix webview-ui
• End-to-end tests in a real VS Code instance: npm run test:e2e (run npm run compile-tests first)
• Real offline inference (opt-in): the test in test/unit/providers/LocalClient.inference.test.ts loads a real GGUF model through LocalClient and generates once. It is skipped unless ORBIT_TEST_MODEL points to a GGUF file, and it needs the built worker, so run npm run compile first.

npm run compile
$env:ORBIT_TEST_MODEL = "$HOME\.orbit\models\Qwen2.5-Coder-7B.gguf"
npx jest test/unit/providers

📄 License

MIT