Lanternkb

What the lanternkb MCP server does

The lanternkb MCP server creates a local index of folders you choose and lets connected assistants search that index by meaning or exact words. It is intended for personal files and project material, including notes, source code, PDFs, Word and OpenDocument files, rich text, saved web pages, and subtitles. Scanned pages and audio can be supported optionally.

Each folder has an access tier. Open folders may provide search results to an assistant, while local-only folders remain unavailable to that assistant. New folders default to local-only. Files containing detected secrets are also treated as local-only, and the detected values are redacted before content is split or embedded.

How it works

The application runs locally as an MCP server over stdio. It does not require a background daemon or listening port, and the README states that it sends nothing over the network apart from model downloads the user approves. Connected clients receive files only from the open folders selected during setup.

Indexing and file reads run in sandboxed worker processes. On macOS, the workers use sandbox-exec; on Linux, they use Landlock and seccomp. A secret scanner checks content before indexing, but it is not a complete security boundary. Values that look like examples or placeholders may remain, and some secrets can be missed.

Search can combine semantic matching with exact-word matching. Optional reranking is available; it is disabled by default on Linux because it runs on the CPU and can take roughly 0.4 to 1 second per search. The project reports benchmark results on a synthetic dataset, not a guarantee for a particular collection.

Setup and configuration

Install with uv on macOS 14 or later on Apple silicon, or on Linux. The documented installation uses Python 3.12 and the embed and ui extras:

uv tool install --managed-python --python 3.12 "lanternkb[embed,ui]"

Run lanternkb setup before connecting an assistant. Setup asks which folders to index, whether an assistant may see them, and which models to use, then builds the index. lanternkb ui opens a settings page with a search box, while lanternkb doctor checks the installation and reports fixes.

For Claude Code, lanternkb mcp-setup prints the registration command and deny rules. For Codex and other MCP clients, run lanternkb serve --tier open as a stdio server and add the equivalent client configuration. The installation uses about 0.9 GB on macOS, plus approximately 1.2 GB for the embedding model.

Tools and capabilities

The lanternkb MCP server supports these documented operations and capabilities:

• Search indexed files using natural-language meaning or exact terms.
• Filter searches by file type and modification age from the command line.
• List files changed recently.
• Search through a local web interface.
• Index supported document and code formats.
• Extend file types, embedding models, and folder rules through plugins.
• Upgrade the installed package with uv.

Limitations and notes

Email, browser history, calendars, and contact cards are not indexed. OCR for scanned PDFs and images is disabled by default. Linux reranking has a higher latency than the default search path. The secret scanner should be treated as a second line of defence: keep credential directories out of indexed locations or mark them local-only.

The server is documented for macOS 14 or later on Apple silicon and Linux. Its access controls protect indexed results, but Claude Code may also have direct file-reading tools; the setup output includes deny rules intended to keep those tools away from local-only folders and the index.