Multi (Nightly) – Frontier AI Coding Agent

Abstracts 30+ AI providers (Claude, Gemini, OpenAI, Anthropic, OpenRouter, Ollama, etc.) behind a unified interface, allowing users to define reusable profiles that bundle provider + model + configuration settings. Profiles persist across sessions and can be switched via UI without reconfiguring API keys or model parameters, enabling seamless provider switching without workflow interruption.

Executes read, write, and edit operations on project files with configurable approval controls. Users can enable auto-approval for file reads, writes, or require explicit confirmation per operation. The agent accesses files within the project scope and can modify code, configuration, and documentation files without manual intervention when approval is granted, enabling hands-off refactoring and code generation workflows.

Allows developers to fork the current agent conversation and task state at any point, creating a parallel branch that preserves the original context while exploring alternative approaches. Forked tasks maintain independent state and can be merged back or abandoned without affecting the original task. This enables safe experimentation with multiple solutions while maintaining a clear audit trail of exploration paths.

Persists agent task state (decomposed subtasks, execution progress, conversational context, intermediate results) to disk or cloud storage, enabling developers to close the IDE and resume work later without losing progress. The 'Restore' feature reconstructs the full task context, including file modifications, shell command history, and agent reasoning, allowing seamless continuation of long-running tasks across multiple sessions.

Analyzes project configuration files (package.json, pyproject.toml, go.mod, Cargo.toml, etc.), build scripts, and dependency manifests to understand the project's tech stack, frameworks, and conventions. The agent uses this understanding to generate code that follows project-specific patterns, uses the correct package manager, respects version constraints, and integrates with existing build/test infrastructure. This ensures generated code is immediately compatible with the project environment.

Accepts deadline constraints as input and uses them to prioritize task decomposition and execution order. The agent estimates task duration based on complexity and available time, reorders subtasks to meet deadlines, and alerts developers if tasks cannot be completed within the specified timeframe. This enables deadline-driven development where the agent adapts its strategy to time constraints.

Monitors developer activity patterns (active file, cursor position, typing speed, pause duration) to understand current focus and work flow. The agent uses this awareness to prioritize relevant suggestions, avoid interrupting deep focus periods, and surface task results at opportune moments. This enables non-intrusive agent assistance that adapts to developer work patterns.

Executes arbitrary shell commands in the host environment with configurable approval gating. Commands run with the same permissions as the VS Code process and can be auto-approved or require explicit confirmation. The agent manages background task execution, allowing long-running processes (tests, builds, deployments) to run asynchronously while the developer continues coding, with task state persisted across IDE sessions via the 'Restore' feature.

Searches the project codebase using semantic understanding to locate relevant code, functions, classes, and documentation. The agent can query the codebase to understand architecture, find dependencies, and locate code patterns without requiring explicit file paths. Search results are ranked by relevance and integrated into the agent's context window, enabling informed code generation and refactoring decisions based on existing codebase patterns.

Breaks down complex coding tasks into actionable sub-tasks, plans execution order, and manages task state across multiple steps. The 'Forking' feature allows branching a task to explore alternative approaches while preserving the original task context, enabling parallel exploration of solutions. Task state is persisted and can be restored, allowing developers to pause, review, and resume work across IDE sessions without losing progress or conversational context.

Extends the agent interface across VS Code, JetBrains IDEs (IntelliJ, PyCharm, WebStorm, GoLand, CLion, RustRover, Android Studio, Rider, PhpStorm, RubyMine), and alternative editors (Cursor, Windsurf, Kiro, Antigravity) through a unified agent architecture. The same agent logic, model configuration, and task state synchronize across platforms, allowing developers to switch editors without reconfiguring the agent or losing task context.

Executes agent tasks in isolated git worktrees or branches, preventing modifications to the current working directory and codebase. The agent creates a temporary branch, performs all file modifications and shell commands within that isolated context, and allows the developer to review changes before merging. This enables safe experimentation and prevents accidental corruption of the main codebase during agent execution.

Fetches and parses web pages to integrate external documentation, API references, and web-based resources into the agent's context. The agent can retrieve HTML content from URLs, extract relevant information, and use it to inform code generation and decision-making. This enables the agent to reference live documentation, API specs, and external resources without requiring developers to manually copy-paste content.

Executes long-running agent tasks asynchronously in the background while the developer continues active coding in the editor. The agent manages task queues, preserves IDE context (open files, cursor position, selections), and allows developers to review task progress or results without interrupting their current work. Background tasks can be monitored via a task panel or sidebar widget and resumed/paused as needed.

Implements configurable approval gates for autonomous agent actions (file reads, writes, shell execution, todo updates), allowing developers to define which operations require explicit confirmation vs. auto-approval. Approval thresholds can be set per operation type, enabling fine-grained control over agent autonomy. The extension tracks approval history and can suggest approval patterns based on past decisions.