atomicmemory/llm-wiki-compiler v0.2.0

on GitHub

First major release since 0.1.1. Ships the complete initial roadmap plus an MCP server for AI agent integration.

Highlights

AI agent integration

• MCP server (llmwiki serve) exposes llmwiki's automated pipelines as Model Context Protocol tools so agents can ingest, compile, query, search, lint, and read pages programmatically. Ships with 7 tools and 5 read-only resources. See the README for Claude Desktop / Cursor setup.

Retrieval and search

• Semantic search via embeddings — pre-filters the wiki index to the top 15 most similar pages before calling the selection LLM. Transparent fallback to full-index selection when no embeddings store exists.
• Provider-specific embedding models: Voyage voyage-3-lite for Anthropic, text-embedding-3-small for OpenAI, nomic-embed-text for Ollama.

Multi-provider support

• Swap LLM backends via LLMWIKI_PROVIDER=anthropic|openai|ollama|minimax.
• Anthropic provider supports ANTHROPIC_AUTH_TOKEN, custom base URLs, and falls back to ~/.claude/settings.json for credentials and model.
• MiniMax provider added via the OpenAI-compatible endpoint.

Wiki quality

• llmwiki lint — six rule-based checks (broken wikilinks, orphaned pages, missing summaries, duplicate concepts, empty pages, broken citations). No LLM calls, no API key required.
• Paragraph-level source attribution — compiled pages now include ^[filename.md] citation markers pointing back to the source files that contributed the content.

Obsidian integration

• LLM-extracted tags (2–4 per concept) and deterministic aliases (slug, conjunction swap, abbreviation) surface in frontmatter.
• Auto-generated wiki/MOC.md groups all concept pages by tag.

Infrastructure

• GitHub Actions CI with Node 18/20/22 build+test matrix plus Fallow codebase health check (required for merges).
• Tests grew from 91 to 211.

Contributors

Thanks to @FrankMa1, @PipDscvr, @goforu, and @socraticblock for their contributions.

Install

```bash
npm install -g llm-wiki-compiler@0.2.0
```