llmwiki turns the transcripts coding agents leave on disk into a searchable, interlinked knowledge base. Adapters read session files from Claude Code, Codex CLI, Cursor, Gemini CLI, Copilot Chat and CLI, and Obsidian, convert new .jsonl files into redacted markdown under raw/sessions/, and a build step compiles a static HTML site with global search, syntax highlighting, dark mode, keyboard shortcuts, an activity heatmap, and a page per session.
The wiki layer follows Andrej Karpathy's LLM Wiki pattern: raw sessions are layer one, and an LLM-written layer of sources/, entities/, concepts/, syntheses/, comparisons/, and questions/ pages linked with wikilinks is layer two. Synthesis backends are configurable, with a local Ollama option and a stub mode, and llmwiki all chains synthesize, build, graph, export, and lint in one command.
Alongside the HTML, every page is published in machine-readable form: .txt and .json siblings at the same URL, plus site-level llms.txt, a flattened llms-full.txt, a schema.org JSON-LD graph, a sitemap, an RSS feed, a robots file, AI navigation instructions, and a build manifest with SHA-256 hashes. An MCP server with twelve tools lets any MCP client query the wiki directly. The tool runs on the Python standard library plus a markdown package, with no server, database, or npm dependency.
Features
- Session adapters: Claude Code and Obsidian in production, with Codex CLI, Cursor, Gemini CLI, and Copilot Chat and CLI in beta; a new agent is one subclass of BaseAdapter plus a fixture and snapshot test
- Static site: a fuzzy-match command palette, client-side syntax highlighting, dark mode, breadcrumbs, session filters, related-page panels, reading times, and print-friendly output
- Machine-readable exports: per-page text and JSON siblings plus llms.txt, llms-full.txt, graph.jsonld, sitemap.xml, rss.xml, robots.txt, and a build manifest
- MCP server: twelve tools covering query, search, list, read, lint, sync, export, confidence, lifecycle, dashboard, entity search, and category browse
- Lint rules: 16 checks spanning frontmatter, link integrity, orphans, freshness, duplicates, index sync, and LLM-assisted contradiction and claim verification
- Confidence and lifecycle: four-factor confidence scoring with decay by content type, and a draft, reviewed, verified, stale, archived state machine with 90-day auto-stale
- Obsidian output: a link-obsidian command that symlinks the project into a vault, plus Dataview dashboards, Templater templates, and tag category pages
- Automation: a session-start hook that syncs in the background, auto-build on sync, and a one-shot pipeline with a strict mode for CI
- Navigation files: generated agent-facing files including CLAUDE.md, AGENTS.md, MEMORY.md, and per-project hot caches, with a consolidation pass that prunes outdated entries
