OrbitCode — AI Pair Programmer

An agentic AI coding assistant for VS Code / Cursor: sidebar chat, streaming output, inline diff, a built-in tool runtime (file I/O, terminal, LSP, REPL, search) and MCP integration.

Just sign in with a LLM Gateway account — models are provided by the gateway, so you don't need to supply API keys for each vendor yourself.

Features

Capability	Description
Multi-model chat	Sidebar chat with the model list served by the gateway's `/v1/models`; Claude / GPT / Gemini / Qwen / DeepSeek / Ollama and more, depending on your account's permissions
Streaming output	Token-by-token rendering, interruptible and retryable
Agentic tool chain	File read/write, terminal commands, REPL, LSP diagnostics/references, web search, skills and subagents — all invoked by the model on its own
Inline diff & checkpoints	A diff is generated before any file is modified and can be Keep or Discard'd; AI edits can also be rolled back in a batch per turn
Plan mode	Read-only toolchain — scout first, then act, so your repo is never modified by accident
Inline edit / explain selection	Edit selected code directly with `Ctrl+I`; explain it with `Ctrl+Alt+E`
Subagents & background tasks	Split long tasks to run in parallel; track progress from the lightbulb panel; scheduled (Cron) tasks supported
Project memory / rules / Skills	Centralize conventions under `.orbitcode/`, loaded on demand by the model to save tokens
MCP support	Connect Model Context Protocol servers; their tools are registered automatically
Codebase index	Local full-text index plus `@codebase` retrieval (optional, no embeddings required)
Bilingual UI	One-click switch between Chinese and English in the sidebar

Installation

Requirement: VS Code / Cursor 1.80+

Option 1: Install from VSIX (recommended)

Obtain orbitcode-<version>.vsix.
VS Code → Extensions (Ctrl+Shift+X) → … in the top right → Install from VSIX… → select the file.
Once the success notification appears in the bottom right, reload the window.

Command-line install:

code --install-extension orbitcode-1.0.0.vsix

Option 2: Install from the Marketplace

If the extension is published to the Marketplace, just search for OrbitCode in the Extensions panel and click Install. If it isn't published yet, use the VSIX method above.

Quick Start

Open the sidebar: click the OrbitCode icon in the activity bar; or press Ctrl+Shift+P → run OrbitCode: Open Chat (you can also use OrbitCode: Open Chat in Editor Panel to open a large panel in the editor area).
Sign in: a sign-in page appears on first use — click Sign in with browser, complete the sign-in in your browser, and it returns to the extension automatically. The status bar and Settings → My Account show the current account.
Pick a model: choose a model from the selector above the input box (the list comes from the gateway). Then just ask a question, or let the AI edit code directly in the conversation.

Type @ in the input box to reference the currently open file and workspace files; @codebase runs a local codebase index search.

Day-to-Day Usage

Chat and context

@ references: typing @ opens a picker (open tabs, workspace files). The selection is attached as context and the host reads its content on demand (size-limited to avoid blowing up the context window).
Add selection/file: Ctrl+Shift+L, or right-click in the Explorer/editor and choose Add to OrbitCode Chat.
Images: paste or drag in a screenshot; multimodal models will read it.
Chat history: the history button in the top bar, or OrbitCode: Show Chat History.

Plan Mode

When enabled, the model may only use read-only tools (read file, search, LSP, ask_user, todo_write, etc.) — ideal for "research first, edit later". Toggle it with the tree icon in the top bar or the command OrbitCode: Toggle Plan Mode; the status bar icon turns into a lock. You can also simply tell the AI "enter plan mode first".

Checkpoints and undo

Before writing files, the AI leaves backups in .orbitcode/checkpoints/, one checkpoint group per user turn. The Rollback button in the top bar reverts all AI edits for a turn in one batch (with a confirmation prompt).

This is a different axis from the Diff Keep / Discard: the diff handles a single file, while checkpoint rollback handles an entire turn of the conversation.

Inline edit / explain selection

Action	Shortcut
Inline-edit the selected code	`Ctrl+I` (macOS `Cmd+I`)
Explain the selected code	`Ctrl+Alt+E`
Keep changes (in diff)	`Ctrl+Enter`
Discard changes (in diff)	`Ctrl+Shift+Enter`

Subagents and background tasks

When the AI tackles a long-running task, it may call spawn_subagent to run work in parallel in the background and push a summary back into the main conversation. The lightbulb panel in the top bar shows recent subagents and scheduled tasks; you can also ask the AI to run list_background_tasks, cron_create or cron_delete to manage them.

Project memory, rules and Skills

Under the project root's .orbitcode/ (legacy .ycode/ and .claudecode/ are also supported):

Path	Purpose
`.orbitcode/memory.md`	Project memory: conventions, test commands, past decisions. Let the AI fill it in with `ingest_project_memory` or append notes with `append_project_memory`
`.orbitcode/rules/`	Project rules, injected into the system prompt
`.orbitcode/skills/*.md`	Skill templates, loaded on demand by the AI via `use_skill`
`.orbitcode/mcp.json`	Project-level MCP server configuration
`.orbitcode/hooks.json`	Lifecycle hooks for before/after tool calls and subagents

MCP (Model Context Protocol)

Add servers (command, arguments, environment variables, enable/disable) under Settings → MCP in the sidebar.
You can also write to the VS Code setting orbitcode.mcpServers, or place .orbitcode/mcp.json in the workspace (governed by orbitcode.enableProjectMcp, on by default and merged).
If local context is tight, disable orbitcode.injectMcpTools to save the per-request schema overhead.

Common Commands

Command	Description
`OrbitCode: Open Chat`	Open the sidebar chat
`OrbitCode: Open Chat in Editor Panel`	Open a large panel in the editor area
`OrbitCode: New Chat` / `Show Chat History`	New / past conversations
`OrbitCode: Focus Chat Input`	Focus the input box (`Ctrl+L`)
`OrbitCode: Inline Edit Selection`	Inline-edit the selection (`Ctrl+I`)
`OrbitCode: Explain Selection`	Explain the selection (`Ctrl+Alt+E`)
`OrbitCode: Toggle Plan Mode`	Toggle plan mode
`OrbitCode: Rebuild Codebase Index`	Rebuild the local codebase index
`OrbitCode: Open Output Log`	View the runtime log (Output panel → OrbitCode)
`OrbitCode: Sign In / Sign Out`	Sign in / sign out

If a shortcut conflicts, search for orbitcode in Keyboard Shortcuts and rebind it.

Settings

Most options can be changed under Settings (gear icon) in the sidebar; you can also search for orbitcode. in VS Code settings.

Setting	Default	Description
`orbitcode.defaultModel`	empty	Gateway model to use; leave empty for the backend default — you can also switch in the model selector
`orbitcode.streaming`	`true`	Streaming output
`orbitcode.contextCompression`	`normal`	Compression strength for long chat history (`relaxed` keeps more; `off` only truncates overly long tool output)
`orbitcode.confirmDangerousTools`	`true`	Require confirmation before running bash / PowerShell / REPL / rename / screenshot, etc.
`orbitcode.preferMcp`	`true`	Prefer MCP tools for external services such as databases
`orbitcode.injectMcpTools`	`true`	Inject MCP tool schemas into every request
`orbitcode.enableProjectMcp`	`true`	Load the workspace's `.orbitcode/mcp.json`
`orbitcode.mcpServers`	`{}`	MCP server configuration
`orbitcode.webSearchProvider`	`google`	Web search: `google` (Serper) / `duckduckgo` / `tavily`; falls back to DuckDuckGo when no key is set
`orbitcode.serperApiKey` / `orbitcode.tavilyApiKey`	empty	Search keys — best entered in the sidebar settings (stored in the local keyring)
`orbitcode.codebaseIndexEnabled`	`false`	Enable the local codebase index and `codebase_search` / `@codebase`
`orbitcode.codebaseIndexWatch`	`true`	Update the index automatically on file changes
`orbitcode.hooksEnabled`	`true`	Run hooks from `.orbitcode/hooks.json`
`orbitcode.subagentMaxParallel` / `MaxTurns` / `TimeoutMinutes`	3 / 25 / 15	Subagent concurrency, turn count and timeout

Data and Privacy

Account credentials: the token issued after sign-in is stored in the VS Code local keyring (ExtensionContext.secrets) and never written in plain text to settings.json.
Search keys (Serper / Tavily): entering them in the sidebar settings also stores them in the keyring; the settings panel offers a "clear stored key" action. If plain-text configuration is detected after upgrading from an older version, it is migrated to the keyring and the plain text is cleared.
Code and conversations: only sent on demand to the selected gateway/model service when making a request. The local codebase index, checkpoints and project memory are all kept inside your own workspace (.orbitcode/).

FAQ

Symptom	What to do
Still shown as signed out after signing in	Run OrbitCode: Sign In with LLM Gateway again to redo the browser callback; make sure the browser can jump back to VS Code
Model list is empty	The gateway must be reachable over the network; retry later or click refresh in the model settings panel
Sidebar is blank / commands not found	Incomplete installation package. Uninstall, then reinstall with the official `.vsix` (do not repackage with options that exclude dependencies)
Want to undo AI edits	Discard a single file in its diff; use Rollback (checkpoints) in the top bar for a whole turn
AI ignores project conventions	Put the conventions in `.orbitcode/rules/` or `.orbitcode/memory.md`
Troubleshooting	OrbitCode: Open Output Log, or select OrbitCode in the Output panel

Feedback

Questions and suggestions: file them via the extension repository's Issues, including the OrbitCode output log and reproduction steps.

License

MIT