Mentat

Mentat analyzes your entire codebase to understand project structure, dependencies, and coding patterns, then uses this context to generate code changes across multiple files simultaneously. It maintains awareness of file relationships and imports, allowing it to generate coherent changes that respect existing architecture rather than isolated snippets. The system indexes relevant files based on user intent and passes them as context to the LLM, enabling context-aware completions that align with project conventions.

Mentat provides a command-line interface where developers can describe coding tasks in natural language and receive streaming code generation responses directly in the terminal. The CLI maintains conversation history within a session, allowing follow-up refinements and iterative code improvement without losing context. It integrates with the user's editor or displays diffs inline, enabling immediate review and acceptance of changes.

Mentat automatically identifies which files are relevant to a coding task by analyzing the user's natural language description and the codebase structure. It uses heuristics like import relationships, file naming patterns, and semantic similarity to prioritize which files should be included in the LLM context. This reduces the need for users to manually specify file paths and ensures the most relevant code context is available for generation.

Mentat generates code changes as unified diffs that users can review before applying them to their codebase. The system shows exactly what will change, allowing developers to accept, reject, or modify individual changes. This prevents blind application of AI-generated code and maintains developer control over the final output. Changes can be applied selectively to specific files or hunks.

Mentat maintains a conversation history within a session that tracks all previous prompts, responses, and accepted changes. This allows users to refine code iteratively by asking follow-up questions or requesting modifications without re-explaining the full context. The system preserves the conversation state, enabling the LLM to understand references to previous changes and build upon them incrementally.

Mentat supports code generation across multiple programming languages (Python, JavaScript, Java, C++, Go, Rust, etc.) while maintaining language-specific syntax and formatting conventions. The system detects the target language from file extensions and project context, then ensures generated code follows the appropriate style and idioms. This enables developers to work with AI assistance regardless of their primary language.

Mentat integrates with Git to understand the codebase history, track which files have been modified, and provide context about recent changes. It can use Git metadata to improve file relevance scoring and understand the project's evolution. Changes generated by Mentat can be automatically staged or committed, and the system is aware of uncommitted changes to avoid conflicts.

Mentat abstracts the underlying LLM provider, allowing users to switch between Claude, GPT-4, local models, or other compatible APIs without changing their workflow. The system handles provider-specific API differences, authentication, and response formatting transparently. Users can configure their preferred provider via configuration files or environment variables.

Mentat supports configuration at multiple levels: user-level defaults (stored in home directory), project-level overrides (stored in the repository), and command-line arguments. This allows teams to enforce consistent settings across a project while letting individual developers customize their experience. Configuration covers LLM provider, model selection, context window size, and other behavior parameters.