Kilo Code: AI Coding Agent, Copilot, and Autocomplete

Converts natural language descriptions into executable code by routing prompts through a configurable AI model (Claude, Gemini, GPT-4, or 500+ alternatives via OpenRouter). Implements a self-verification loop where the generated code is re-evaluated by the same or different model to check correctness before insertion into the editor. Supports multi-turn refinement where users can iteratively request changes to generated code within the same context window.

Provides context-aware code completion suggestions as the user types, triggered automatically or on-demand via keybinding. Integrates with VS Code's InlineCompletionItemProvider API to display suggestions inline without interrupting editor flow. Streams completions from selected AI model (Claude, GPT-4, Gemini, or 500+ alternatives) with latency optimized for real-time interaction. Respects user's current file context, language syntax, and project structure to generate relevant suggestions.

Implements transparent pricing model where users are charged at provider rates (OpenAI, Anthropic, Google) without markup. Billing is aggregated through OpenRouter or direct provider APIs, and users see per-token costs for each request. No subscription required — users pay only for tokens consumed. Pricing is displayed in UI before requests are sent, enabling users to make informed decisions about model selection.

Manages API keys for model providers (OpenAI, Anthropic, Google, OpenRouter) with optional user account creation. Users can provide their own API keys (stored locally in VS Code settings) or create a Kilo Code account to access shared API keys. Account creation is optional — users can use the extension without creating an account if they provide their own API keys. Key storage mechanism is undocumented but likely uses VS Code's SecretStorage API for encryption.

Supports code generation and completion across 40+ programming languages (JavaScript, Python, TypeScript, Go, Rust, Java, C++, C#, PHP, Ruby, Kotlin, Swift, etc.). Language detection is automatic based on file extension, and the AI model is prompted with language-specific context (syntax, idioms, frameworks). Completion suggestions respect language-specific conventions (e.g., snake_case for Python, camelCase for JavaScript). No language-specific tuning is documented — same model is used for all languages.

Applies AI-driven refactoring transformations to selected code, entire files, or project-wide patterns. User specifies refactoring intent (e.g., 'extract method', 'rename variables for clarity', 'convert to async/await') in natural language, and the selected model generates refactored code while preserving functionality. Integrates with VS Code's edit API to apply changes atomically, with undo support. Scope (selection, file, or project) is user-controlled via command palette or sidebar UI.

Generates shell commands from natural language descriptions (e.g., 'find all TypeScript files modified in the last week') and executes them in the user's local terminal with explicit user confirmation. Integrates with VS Code's terminal API to run commands in the integrated terminal, capturing output and displaying results in the editor or terminal panel. Supports bash, zsh, PowerShell, and other shells based on user's environment.

Automates browser interactions (clicking, typing, navigation, form submission) based on natural language instructions. Mechanism is undocumented but likely uses Playwright or Puppeteer to control a browser instance. User describes desired browser action (e.g., 'fill in the login form and submit'), and the AI generates automation code or directly controls the browser. Supports multi-step workflows (e.g., navigate to URL, fill form, submit, verify result).

Abstracts away model provider differences by routing requests to 500+ AI models (Claude, GPT-4, Gemini, Llama, Mistral, etc.) via OpenRouter API. Users select a model via UI dropdown or configuration, and all requests (code generation, completion, refactoring) are routed to that model without extension changes. Supports switching models mid-session and per-task model selection (e.g., use Claude for planning, GPT-4 for coding). Handles API key management, rate limiting, and billing aggregation via OpenRouter.

Discovers and integrates Model Context Protocol (MCP) servers to extend agent capabilities beyond built-in features. Users can install MCP servers from a marketplace or configure custom servers, and the agent automatically discovers available tools (e.g., file system access, database queries, API calls). Requests are routed through MCP protocol to appropriate servers, enabling the agent to perform complex multi-step tasks (e.g., 'query the database and generate a report'). Integration is transparent — users describe intent in natural language, and the agent selects appropriate MCP tools.

Enables users to create custom agent modes (specialized workflows) tailored to specific tasks. Pre-built modes include 'Architect' (planning), 'Coder' (code generation), and 'Debugger' (debugging). Users define custom modes by specifying system prompts, tool availability, model preferences, and execution constraints. Modes are invoked via command palette or sidebar, and the agent operates within the mode's constraints (e.g., 'Architect' mode may prioritize planning over code generation). Mode configuration is stored locally in VS Code settings.

Enhances code completion by analyzing project structure, imports, and dependencies to provide contextually relevant suggestions. Completion engine infers user's intent from cursor position, recent edits, and file structure (e.g., if user is in a React component, suggestions favor React patterns). Integrates with language servers (LSP) to understand syntax and type information, enabling more accurate suggestions than simple pattern matching. Context is sent to selected AI model along with completion request.

Analyzes runtime errors, stack traces, and test failures to suggest fixes. User selects error output or stack trace, and the 'Debugger' mode analyzes the error using the selected AI model to identify root cause and suggest code changes. Integration with VS Code's debug adapter protocol (DAP) enables breakpoint inspection and variable analysis. Suggestions include code patches, configuration changes, or debugging strategies.