Continue - open-source AI code agent

Provides real-time code suggestions as developers type within VS Code editor, leveraging the current file context and potentially project-level code patterns. The autocomplete feature integrates directly into VS Code's IntelliSense pipeline, intercepting typing events and returning LLM-generated completions that appear alongside traditional language server suggestions. Completion requests are sent to configured AI models (Claude, GPT-4, or others) with the current file buffer and cursor position as context.

Enables developers to select code regions and request AI-driven modifications (refactoring, bug fixes, style changes) that are applied directly to the editor without leaving the current file. The Edit feature sends the selected code snippet plus surrounding context (file header, imports, function signatures) to the configured LLM, receives a transformed version, and displays a diff preview before applying changes. This pattern avoids context loss and allows iterative refinement within the same editing session.

Implements error handling and fallback mechanisms when primary LLM requests fail due to API errors, rate limits, or network issues. The system can automatically retry failed requests, switch to a fallback model, or degrade gracefully by disabling features temporarily. Error messages are user-friendly and suggest remediation steps (e.g., check API key, wait for rate limit reset).

Respects VS Code's workspace trust settings and only enables Continue features in trusted workspaces, preventing accidental code exposure in untrusted projects. The system integrates with VS Code's native workspace trust API to determine trust status and can restrict file access, API calls, and code generation based on trust level. This prevents malicious code or untrusted dependencies from being analyzed by Continue.

Allows developers to define project-specific Continue settings in a `.continue` directory or configuration file at the project root, enabling team-wide customization of model selection, context injection, and feature behavior. Configuration is version-controlled alongside code, ensuring consistency across team members and CI/CD environments. Settings can override global Continue configuration for specific projects.

Provides a sidebar chat interface where developers can ask questions about code, request explanations of specific functions or files, and receive natural language responses from the configured LLM. The Chat feature maintains conversation history within a session, allows developers to reference code snippets or files by selection, and can answer both general programming questions and project-specific queries. Context is built from the current file, selected text, and optionally the broader project structure depending on configuration.

Enables developers to assign high-level development tasks (e.g., 'add unit tests for the auth module', 'refactor this component to use hooks') to an AI agent that breaks down the task into steps, executes code modifications, and reports progress within VS Code. The Agent feature uses chain-of-thought reasoning to plan task decomposition, iteratively generates and applies code changes, and can reference the codebase to understand dependencies and context. This differs from one-off edits by maintaining task state across multiple LLM calls and file modifications.

Allows developers to configure and switch between multiple LLM providers (OpenAI, Anthropic, Mistral, local models via Ollama or LM Studio) within a single VS Code session. The configuration system supports per-feature model assignment (e.g., use GPT-4 for Agent tasks, Claude for Chat), API key management, and custom endpoint configuration for self-hosted or on-premise LLM deployments. Model switching is seamless and does not require extension reload.

Automatically indexes the developer's codebase to provide semantic context to LLM requests, enabling the AI to understand project structure, dependencies, and coding patterns without explicit file selection. The context system analyzes imports, function definitions, and class hierarchies to determine relevant code snippets for each request, then includes them in the LLM prompt. This reduces the need for manual context specification and improves the relevance of AI-generated code.

Enables developers to define AI-driven code checks (linting, style validation, security analysis) that run in CI/CD pipelines and enforce AI-generated rules as part of the build process. Checks are defined in configuration files (YAML or JSON) and can be version-controlled alongside code, ensuring consistency across the team. The integration allows AI to analyze pull requests, suggest improvements, and block merges if violations are detected.

Allows developers to customize keyboard shortcuts for each Continue feature (Autocomplete, Edit, Chat, Agent) within VS Code's keybinding configuration system. Shortcuts can be context-aware (e.g., only active when code is selected for Edit feature) and can be bound to custom commands or predefined Continue actions. This enables developers to integrate Continue into their existing workflow without learning new keybindings.

Generates code that respects language-specific syntax, conventions, and idioms by analyzing the current file's language and applying language-specific templates or patterns. The system detects the file extension or language mode and adjusts LLM prompts to include language-specific context (e.g., Python indentation conventions, JavaScript async/await patterns). This improves code quality and reduces syntax errors in generated code.

Renders LLM responses progressively as tokens arrive from the inference endpoint, displaying partial results in real-time rather than waiting for the complete response. This pattern is implemented in Chat and Edit features, allowing developers to see AI reasoning and code generation as it happens. Streaming reduces perceived latency and enables developers to interrupt long-running requests if the output is clearly incorrect.