What Happened
A developer released Tessera, an open-source tool that gives Claude Desktop the ability to search across your entire local workspace and remember context between sessions. It connects through Anthropic's Model Context Protocol (MCP), so Claude can call it automatically during conversations without manual file uploads.
The setup is straightforward: pip install, then tessera init for interactive configuration. No Ollama, no Docker, no API keys required. It downloads a roughly 220 MB embedding model on first run, then indexes your designated folders. Supported file types include Markdown, CSV, and session logs.
Under the hood, Tessera uses fastembed for ONNX-based local embeddings and LanceDB as the vector store. Documents get chunked at approximately 800 characters per segment. Search is hybrid, combining semantic similarity with keyword matching for better recall. Subsequent syncs only re-index changed files, so the overhead stays low.
Claude Desktop gets access to four MCP tools: search_documents for querying your files, remember and recall for persistent memory, and knowledge_graph for visualizing how your documents connect to each other. Everything runs locally. No data leaves your machine.
Why It Matters
Claude Desktop's biggest limitation has always been context isolation. You can attach a file, but Claude cannot search across a folder of 200 Markdown notes to find the one that answers your question. Every new conversation starts from zero. Tessera solves both problems at once.
For anyone maintaining project documentation, research notes, or decision logs, this changes how you interact with Claude. Instead of copying and pasting context into every conversation, you point Tessera at your docs folder and Claude just knows. Ask "what did we decide about the API design?" and it pulls the relevant meeting notes automatically.
The MCP integration means this works transparently. You do not have to learn a new interface or switch tools. Claude calls Tessera's search when it needs context, the same way it would call any other MCP tool. The remember and recall functions act as a persistent scratchpad, so important decisions survive across sessions.
Our Take
This fills a real gap. If you have been using Claude Desktop daily, you know the lack of workspace awareness is the single most frustrating limitation. Every session is an amnesiac starting over. Tessera is not the first attempt to fix this, but the zero-dependency install and MCP-native approach make it the most practical solution I have seen.
The choice of LanceDB and fastembed is sensible. LanceDB is fast, file-based, and does not need a running server. Fastembed handles embeddings locally through ONNX, so there is no OpenAI API cost or network dependency. The 800-character chunking is on the small side - you may get fragmented results from longer documents - but it is a reasonable default for notes and logs.
The knowledge graph visualization is a nice addition, though I suspect most users will rely on search_documents and remember for 90% of their usage. The real test will be how well it handles larger workspaces with hundreds of files, mixed formats, and frequent updates. The incremental sync is promising, but I would want to see it proven at scale.
If you use Claude Desktop as your primary AI interface and keep any kind of local documentation, Tessera is worth trying today. The install is five minutes, the risk is zero, and the upside is a Claude that actually knows your project.