hashcode is a pioneering desktop application designed to enhance the user experience of interacting with large language models (LLMs) such as OpenAI's GPT and Anthropic's Claude. Unlike standard AI chat applications that only allow users to send and receive messages, hashcode empowers users by providing visibility and control over the context—essentially the black box that the AI utilizes to generate responses.

Key Features

Context Map

hashcode presents a comprehensive Context Map that visualizes the entire context a model consumes:

• Numbered Nodes: Each user/assistant interaction is represented as a distinct node, allowing for easy tracking.
• Token Weight Colors: Instantly identify problematic areas with color-coded token weights—green for normal, yellow for heavy usage, and red for very heavy usage.
• Minimap: Provides a bird’s-eye view of the context for quicker navigation.

AI Editing Interface

The Context Workbench allows users to edit AI's context seamlessly, featuring:

• Suggestions: Automatically identifies bloated nodes and redundant outputs.
• Manual Edits: Users can chat with the context model to precisely delete or compress nodes.
• Version Control: Enables users to browse and restore previous context revisions effortlessly.

Precision Editing Tools

hashcode offers advanced tools for granular control over context items, including:

get_node_details: Inspect detailed contents of a node.
delete_item: Remove specific elements from any node.
replace_item: Substitute existing content with revised information.
compress_item: AI-compress items, maintaining their type.
compress_nodes: Summarize multiple nodes into a cohesive node.
delete_nodes: Remove entire nodes that are no longer relevant.

Version Control for Context

Every edit is versioned, allowing users to revert to previous states easily. This snapshot feature ensures that users can keep their context organized without fear of permanent loss:

Revision #1 ← "Compressed weather tool outputs" [Restore]
Revision #2 ← "Deleted redundant shell commands" [Restore] ← Active
Revision #3 ← "Merged nodes #2-5 into summary" [Restore]

Multi-Provider Support

hashcode is designed to connect with multiple LLMs, providing flexibility in model choice and usage:

• OpenAI, Claude, and Gemini are all supported with straightforward configuration.

User-Friendly Desktop Client

Built on Electron, the application acts as a seamless desktop experience, featuring:

• A clean three-panel layout for efficient user interaction.
• Dark theme to reduce eye strain during prolonged use.
• Support for markdown rendering, enhancing text formatting capabilities.

Architecture

The application operates on a unique Two-Model Architecture, which separates the main AI model from the context management model. This structure allows for flexible editing and prevents conflicting interactions during edits:

Main Model ← Chat Interactions
Context Model ← Context Inspection & Editing

Conclusion

hashcode represents a step forward in AI interaction, offering tools that make the context visible, editable, and versioned, thereby granting users unprecedented control over their interactions with LLMs. Contributions, feedback, and testing are encouraged as this project is in its early alpha phase. Aiming to redefine how users engage with AI, hashcode is set to innovate the landscape of AI development.