ThreadShelf Archives and Searches AI Conversations Locally

Summary

ThreadShelf is a single-user, local-first workspace for collecting AI conversations from ChatGPT, Claude, Google AI Studio, OpenRouter, LM Studio, and Grok. Users point it at exported or locally stored files; provider-specific parsers normalize conversations, create local multilingual embeddings, and store vectors and thread snapshots in LanceDB. The web interface supports semantic search, exact substring search, role filters, complete thread reconstruction, Markdown or JSON export, saved searches, pins, and archive insights. The same index is available through a command-line interface, HTTP API, and MCP stdio server. Archived conversations remain searchable after indexed source files are moved, rewritten, or deleted, although older collections may require re-indexing to gain that persistence behavior. ThreadShelf also offers an Experimental Beta continuation layer: managed llama.cpp runs a selected GGUF model through a loopback-only server, while an explicitly marked OpenRouter provider sends selected conversation context and prompts off-device. Archived thinking turns are excluded from generation, and local private conversations are kept out of the thread store and semantic index. A guided setup can resolve and display the llama.cpp build and model download plan, including URLs, SHA-256 digests, sizes, and destinations, before confirmation. Hugging Face browsing and downloads are supported for GGUF models, with file verification and hardware-fit assessments. The project requires Node.js 20.19+ and npm, and warns that several provider export formats are undocumented and may change. Its HTTP API has no authentication, so the documentation recommends loopback or deliberate trusted-network configuration.