A new open-source project called Luma is redefining how AI agents interact with local file systems by exposing a shared semantic index directly to them. While traditional Linux launchers focus on human navigation, Luma treats the agent as a first-class user, allowing it to query photo content and text-in-image data without walking the disk. The project, hosted on GitHub by sakurartro, leverages CLIP for semantic search and PaddleOCR for text extraction, creating a unified data layer that serves both the human interface and the machine logic.

The Agent Advantage

The core value proposition lies in speed and efficiency for autonomous workflows. According to the project documentation, a standard agent task to find a specific photo takes 1 minute and 26 seconds when using traditional file walking methods. With Luma’s pre-built index, the same query on Claude Code using Sonnet 5.5 at medium effort completes in just 11 seconds. This massive reduction in latency is achieved because the agent doesn't need to open files one by one; it simply asks the launcher's MCP server for matches, receiving results in under a second.

Technical Architecture and Integration

Luma implements this capability through a Model Context Protocol (MCP) server, accessible via luma mcp, and a CLI interface via luma query. The server exposes two primary tools: luma_search_images for visual queries and luma_search_text for OCR-based searches. Integration is streamlined for popular coding agents, with one-click connection support for Claude Code, Codex, and opencode. For manual setup, users can register the server using claude mcp add luma --scope user -- luma mcp, provided the agent_api flag is enabled in the local configuration.

Privacy and Local-First Design

Critically, Luma maintains a strict local-first posture. The CLIP and OCR models, which require approximately 154 MB and 12 MB downloads respectively, run entirely on the user's machine. The index, thumbnails, and agent queries never leave the device, addressing common privacy concerns associated with cloud-based AI indexing. The only external network calls are optional model downloads and cached currency exchange rates, ensuring that the agent’s context remains private and contained within the user's Linux environment.

Key Takeaways

  • Luma reduces AI agent photo search latency from ~86 seconds to ~11 seconds by using a pre-indexed CLIP/OCR database.
  • The tool exposes luma_search_images and luma_search_text via an MCP server for seamless integration with Claude Code and Codex.
  • All indexing and model inference happen locally, ensuring no user photos or query data leave the machine.
  • The launcher supports major Linux desktop environments including Hyprland, niri, GNOME, and KDE, with specific keybind and window rule configurations.

The Bottom Line

This isn't just another launcher; it's a middleware layer that makes local AI agents actually useful for desktop automation. By decoupling the expensive indexing process from the query time, Luma solves the biggest bottleneck for agents trying to interact with personal data.