OpenClaude VS Code

Provides a VS Code sidebar chat panel that streams responses from 8+ LLM providers (OpenAI, Anthropic, Google Gemini, Ollama, AWS Bedrock, GitHub Models, and OpenAI-compatible custom endpoints) with runtime provider switching via `/provider` slash command or UI badge. The extension wraps the OpenClaude CLI, delegating model inference to the CLI process while rendering markdown-formatted streaming responses with syntax-highlighted code blocks in the native VS Code chat interface. Provider credentials are configured via environment variables (OPENAI_API_KEY, GOOGLE_API_KEY, etc.) or interactive setup commands.

Implements a @-mention system (similar to Slack or GitHub) allowing developers to explicitly include file contents, entire folders, or specific line ranges in chat context without automatic project-wide scanning. When a user types `@filename.js`, `@folder/`, or `@file.js:10-20`, the extension resolves the path relative to the workspace root, reads the file contents, and injects them into the LLM context window. This approach avoids token waste on irrelevant files and gives developers fine-grained control over context scope, critical for large codebases where full project indexing would exceed token limits.

The extension integrates with the Model Context Protocol (MCP), an open standard for extending LLM context with external data sources and tools. The extension includes an MCP plugin manager that allows developers to install and configure MCP servers (e.g., for accessing databases, APIs, file systems, or custom knowledge bases). When an MCP server is enabled, the extension automatically includes its resources and tools in the LLM's context, allowing the AI to query external data sources or invoke external tools. This architecture decouples context sources from the extension itself, enabling extensibility without modifying the extension code.

The extension includes an interactive onboarding walkthrough that guides new users through initial setup, including provider selection, API key configuration, keybinding explanation, and feature overview. The walkthrough is likely triggered on first installation and can be re-triggered via a command. It provides a structured, step-by-step introduction to the extension's capabilities, reducing the learning curve and setup friction. The walkthrough may include interactive examples (e.g., 'try asking the AI a question') to familiarize users with the chat interface.

The extension provides global keyboard shortcuts for common actions: `Cmd+Escape` (Mac) / `Ctrl+Escape` (Windows/Linux) to open/focus the chat panel, and `Cmd+Shift+Escape` / `Ctrl+Shift+Escape` to open the chat in a new tab. Additionally, `Alt+[key]` shortcuts enable quick @-mention insertion (exact keys not fully documented). These shortcuts are registered with VS Code's keybinding system and can be customized by users via the keybindings.json file. The shortcuts provide quick access without requiring mouse navigation or command palette usage.

When the LLM generates code changes, the extension renders them in VS Code's native diff viewer (side-by-side or unified diff format), allowing developers to review proposed edits before applying them. The workflow is: AI generates code → extension parses response for code blocks → creates a temporary file or diff representation → opens native VS Code diff UI → developer clicks 'Accept' (applies changes) or 'Reject' (discards). This integrates seamlessly with VS Code's built-in diff viewer, avoiding custom UI and leveraging familiar editor affordances.

The extension maintains a persistent conversation history for each chat session, allowing developers to browse past conversations, resume interrupted sessions, and fork conversations at any point to explore alternative paths. The architecture stores conversation metadata (messages, model used, provider, timestamp) locally or in extension storage, enabling quick retrieval without re-querying the LLM. Forking creates a branch point in the conversation tree, allowing developers to ask 'what if' questions without losing the original conversation thread. This is similar to ChatGPT's conversation management but integrated into VS Code's sidebar.

As the LLM generates tokens, the extension streams them to the VS Code chat panel in real-time, parsing markdown syntax and rendering code blocks with language-specific syntax highlighting. The implementation uses a markdown parser (likely a lightweight library) to identify code fences (triple backticks with language specifiers), extract the language identifier, and apply VS Code's built-in syntax highlighter for that language. Streaming is non-blocking — the UI updates incrementally as tokens arrive, providing immediate feedback to the developer. The extension also supports interrupting the stream via a 'Stop' button.

When the LLM calls external tools (e.g., file system operations, API calls, code execution), the extension renders these invocations as collapsible blocks in the chat panel, showing the tool name, input parameters, and execution result. The implementation parses tool-calling responses from the LLM (likely using OpenAI's function-calling format or similar), executes the tool via the OpenClaude CLI or native handlers, and streams the result back into the chat context. This provides transparency into what the AI is doing beyond just generating text — developers can see file reads, API calls, and other side effects.

The extension displays a live token count and estimated cost in the VS Code status bar, updating as the user types messages and the LLM generates responses. The implementation likely tokenizes the current chat context using the selected LLM's tokenizer (e.g., OpenAI's cl100k_base for GPT-4, Anthropic's tokenizer for Claude) and multiplies by the provider's per-token pricing to estimate cost. This helps developers stay aware of API spending and avoid expensive queries. The status bar widget is clickable, potentially opening a cost breakdown or provider settings.

The extension integrates with Git worktrees, allowing developers to run separate OpenClaude sessions for different branches or worktrees without context collision. When a developer switches to a different Git worktree (e.g., `git worktree add ../feature-branch`), the extension detects the change and isolates the chat session, conversation history, and context to that worktree. This enables parallel development workflows where multiple team members or the same developer can work on different branches with independent AI assistance without mixing context.

The extension provides 5 permission modes (Default, Plan, Accept Edits, Bypass, Don't Ask) that control how the AI can modify files and what approval is required. While the exact behavior of each mode is not fully documented, the architecture likely works as follows: each mode defines a threshold for automatic approval (e.g., 'Bypass' auto-applies all edits, 'Accept Edits' requires explicit approval, 'Don't Ask' disables edits entirely). The extension checks the current permission mode before applying AI-suggested changes and enforces the corresponding workflow. This provides fine-grained control over AI autonomy, from fully autonomous to fully manual.

The extension provides a `/provider` slash command that opens an interactive setup wizard for configuring LLM providers without manually editing environment variables. The wizard likely walks through provider selection (OpenAI, Anthropic, Gemini, etc.), API key entry, model selection, and optional advanced settings (base URL for custom endpoints, temperature, max tokens). The configuration is stored in VS Code's extension settings or a local config file, and the extension reads these settings at startup. This lowers the barrier to entry for non-technical users and reduces configuration errors from typos in environment variables.