Graft Solves AI Agent's Contextual Challenges

This means it can integrate directly with existing AI agent subscriptions without additional overhead. Users can also leverage Graft for optional document generation, aiding in code explanation and understanding.

The tool also reduced agent tool usage by 46% and token consumption by 42%. These efficiencies collectively led to an average overall cost reduction of 32% for agent operations.

Instead of converting code into numerical representations and matching by similarity, Graft reads the code directly to identify actual dependency relationships. This method allows for incremental updates without requiring full rebuilds of the entire project map, maintaining efficiency as code evolves.

This structured map is stored as a JSON file, providing a clear and machine-readable representation of the codebase. A browser-based viewer is also available, enabling developers to visually explore the intricate web of code dependencies.

With the CLI setup, Graft automatically guesses relevant locations from a user's prompt and attaches them to every message, regardless of whether the agent specifically requested them. Conversely, the MCP approach allows agents to query Graft only when necessary, providing more control and potentially higher accuracy, as internal tests have shown.

This process involves checking for code changes since the last map build and, if detected, updating the map without involving the AI model. The system also offers simple installation through provided setup commands, streamlining its deployment within development environments.

To utilize Graft within a project, users must run the `init` command in the project's terminal. This command is crucial as it adds project-specific instructions directly into the working folder, which would be lost if executed elsewhere. During this setup phase, users are also required to select their specific coding agent, tailoring Graft's integration to their development environment.

These hooks provide initial instructions to the model at the start of a session, attach relevant file locations to prompts, and ensure the map is updated after any code modifications. This mechanism guarantees that the agent consistently operates within the defined graph workflow, maintaining efficiency and accuracy.

The `graft build` command is used to generate the initial map for existing projects. For new, empty folders, the system automatically builds the map as files are created and added. Subsequently, hooks ensure that the map remains up-to-date after every file modification, allowing users to run Claude Code normally while the map viewer continuously displays the updated nodes and edges created by Graft.

Performance tests involving the development of a scheduling application with Fable 5.1 demonstrated Graft's impact. Building the app with Graft took 39 minutes and utilized approximately 31% of the context window. Without Graft, the same build extended to 47 minutes and consumed around 35% of the context window. Both applications achieved virtually identical functionality, highlighting Graft's efficiency gains even during initial builds. These gains become even more pronounced during subsequent iterative changes, showcasing its value in ongoing development.

This means that non-code contexts, such as Product Requirements Documents (PRDs), learning notes, or general project documentation, are excluded from the graph. When an AI agent needs to interact with these non-code files, it reverts to its usual default methods, as Graft's mapping capabilities are strictly limited to code structures and dependencies.