lore

README

lore

A local-first documentation index served to LLM agents over MCP. Single static Go binary — no Node, no Python, no CGO, no hosted service required. lore ingests docs from GitHub repos, llms.txt sites, package registries, and crawled websites, indexes them with full-text (and optionally semantic) search, and serves them to any MCP client.

Comparable to context7, docs-mcp-server, rtfmbro-mcp, and contextmine — see issue #1 for the full design spec and the reasoning behind every architectural decision below.

Install

go install github.com/cheetahbyte/lore/cmd/lore@latest

Prebuilt cross-platform binaries are published to GitHub Releases on tag push.

Usage

lore add github:owner/repo        # index a GitHub repo's docs (latest tag, or default branch)
lore add npm:react                # index an npm package's readme
lore add pypi:requests             # index a PyPI project's description
lore add pkg.go.dev:golang.org/x/tools
lore add llms-txt:example.com      # index a site's llms.txt / llms-full.txt
lore add url:https://docs.example.com --depth 2 --include /docs/

lore list                          # what's indexed
lore search github:owner/repo "how do I configure X"
  lore refresh                       # re-fetch everything, skip unchanged content
  lore sync                          # discover and index dependencies from the current project
  lore watch                         # keep dependency docs current while manifests change
  lore serve                         # run the MCP server over stdio

Point any MCP client at lore serve. It exposes three tools: list_libraries, resolve_library, search_docs.

lore sync discovers dependencies from package.json/package-lock.json, go.mod, and requirements.txt, preferring exact versions from lockfiles. lore watch runs the same sync when those manifests change. Bare references such as lore add react are also inferred against supported package sources.

Config

Global, not per-project — one index at $XDG_DATA_HOME/lore/index.db (~/.local/share/lore by default), one config at $XDG_CONFIG_HOME/lore/config.toml:

[embeddings]
provider = ""   # "" = FTS-only (default), or "openai" | "ollama" to enable vector search
api_key = ""    # or set LORE_EMBEDDINGS_API_KEY
endpoint = ""   # e.g. http://localhost:11434/v1 for Ollama

[sources."url:https://docs.example.com"]
depth = 2
include = ["/docs/"]

Status

Implements the spec in issue #1 end to end: all four source adapters, structure-aware chunking, FTS5 + optional vec1-fused-via-RRF retrieval, the three MCP tools, and the full CLI. Known gaps, tracked as follow-ups rather than blocking:

• Bare-name source inference (lore add react without a npm: prefix) isn't implemented — the type prefix is required for now.
• lore refresh always does a full re-fetch + content-hash diff; the conditional-request optimization (ETags/commit SHAs) from issue #12 isn't wired up yet, so refreshing unchanged content still costs a fetch, just not a re-index.
• The retrieval-quality fixture corpus and CI eval from issue #11 aren't set up.