Augment Code

Analyzes user requests against the entire codebase using semantic filtering (reducing 4,456+ sources to 682 relevant ones) and generates numbered, actionable task lists before any code execution. Users can add, skip, or modify steps before the agent proceeds. This plan-first approach enables structured multi-file changes while maintaining human oversight at the decision point, not just execution point.

Executes planned tasks sequentially while creating checkpoints at each step, allowing users to accept changes, revert to any prior checkpoint, or redirect the agent mid-task without losing work. Each checkpoint captures file state and execution context, enabling granular rollback without manual version control. Integrates with Git for version tracking but provides finer-grained undo than traditional commits.

Allows users to create and maintain workspace Rules — persistent, user-approved memory items that capture project-specific patterns, conventions, and decisions. Rules are stored in the workspace and applied across all agent sessions, enabling the agent to learn from user feedback without automatic memory accumulation. Users explicitly approve, edit, or discard each memory before it's saved.

Uses a credit-based consumption model where tasks consume credits based on complexity and resource usage. Credits are purchased in tiers (Indie: 40k/month, Standard: 130k/month, Max: 450k/month) with auto top-up at $15 per 24k credits. Credits are consumed by agent execution and code review tasks. The exact credit-to-token mapping and per-task cost estimation are not published.

Provides native plugins for VS Code and JetBrains IDEs (IntelliJ, PyCharm, etc.) that embed the agent directly into the development environment. Users interact with the agent through IDE UI elements (sidebar, inline suggestions, context menus) without leaving their editor. The plugin architecture maintains local IDE state while communicating with the cloud-hosted agent.

Provides Augment CLI, a terminal-based interface to the agent that uses the same Context Engine and planning logic as the IDE plugins. Enables developers who prefer terminal workflows to use the agent without opening an IDE. CLI supports piping, scripting, and CI/CD integration.

Provides enterprise-grade security features including SOC 2 Type II compliance, CMEK (Customer-Managed Encryption Keys), ISO 42001 compliance, SIEM integration, data residency options, granular access controls, comprehensive audit trails, and enterprise SSO (OIDC, SCIM). These features are available on Enterprise tier and ensure data protection, regulatory compliance, and organizational control.

Maintains a 'live understanding' of the entire codebase by indexing code, dependencies, architecture, and history, then performs semantic filtering to surface only relevant context (reducing 4,456+ sources to 682 relevant ones per example). Uses a proprietary Context Engine to determine relevance without exposing the filtering mechanism. Stores user-approved memories as workspace Rules that persist across sessions.

Supports multiple LLM backends (Claude Opus 4.5, Opus 4.6, Gemini 3.1 Pro) with user-selectable model configuration. The agent's planning and execution logic is model-agnostic, allowing users to choose models based on task complexity, cost, or latency requirements. Model selection mechanism and whether users can bring custom models remain undisclosed.

Executes arbitrary terminal commands and invokes external tools (e.g., linters, test runners, build systems) as part of task execution. The agent can run commands, capture output, and use results to inform subsequent steps. Integrates with MCP (Model Context Protocol) for custom tool definitions, allowing teams to extend the agent with domain-specific tools.

Integrates with GitHub to read repository context, create pull requests, and provide inline code review comments. Supports multi-organization repositories, allowing the agent to work across different GitHub orgs without reconfiguration. PR summaries and inline comments are generated by the agent, providing context-aware review feedback.

Sends task completion notifications, status updates, and approval requests to Slack channels. Enables asynchronous team collaboration by surfacing agent activity in team communication channels. Available on Standard tier and above, allowing teams to stay informed without constant IDE monitoring.

Implements complete features (e.g., OAuth flows, JWT token rotation) across multiple files while maintaining awareness of dependencies and cross-file references. The agent reads existing patterns, creates necessary files, modifies related code, and ensures consistency across the codebase. Uses codebase-aware context to understand existing architecture and apply similar patterns.

Identifies and fixes bugs by analyzing error messages, stack traces, and test failures, then validates fixes by running tests. The agent can execute terminal commands to run test suites, interpret results, and iterate on fixes. Uses codebase context to understand the bug's scope and potential side effects.

Refactors code across multiple files while preserving consistency with existing patterns and architecture. The agent understands the codebase structure, identifies refactoring opportunities, and applies changes consistently across all affected files. Uses semantic context to ensure refactored code aligns with project conventions.