LLM wiki starter kit (wiki)

Start and run an LLM-maintained wiki (a folder of Markdown notes that an agent keeps: an Obsidian vault, a research notebook, a wiki about a codebase) with any agent that can run shell commands: Claude Code, Codex, Cursor, Copilot, OpenCode and others.

• wiki new sets up a wiki from a preset: folder layout, page templates, an AGENTS.md telling the agent how to be the librarian, and a config whose relation rules match the templates. Presets: research (sources in, pages out), code (a wiki in its own repo describing a code repo) and writing (stories you write in Obsidian, and a wiki of their characters, places and plotlines).
• wiki guide prints the workflow an agent follows (ingest a source, answer a question, audit the wiki, sync after code changes), versioned with the command so the steps never go stale.
• Search and navigation find the best starting page without reading an index, show where each page leads through typed relations with reasons, and read only the sections needed, within a page limit.

How it fits together: you ask your agent to add a source or answer a question; the agent follows the steps wiki guide prints, and uses wiki search, wiki nav and wiki check as it writes the pages. The wiki command never writes a page's content, calls no AI model, and needs no network after a one-time model download. What it builds for search lives in a disposable cache inside the wiki.

New here? Start with docs/getting-started.md, then docs/workflows.md for what you do with a wiki, step by step: asking questions, finding gaps, a code task from plan to commit, researching a topic, a writing session.

More (all the docs):

• docs/concepts.md: how the pieces fit, and what each term means
• docs/presets.md: what each preset is for, and its page types
• docs/guides.md: the guides agents follow, and how to customise them
• docs/commands.md: every command and option
• docs/config.md: every config setting
• docs/troubleshooting.md: what a message means, and what to do

Install

Requires uv, which installs Python 3.13 for the tool if needed.

uv tool install llm-wiki-kit    # once per machine; the command is `wiki`
wiki models download            # once: about 0.2 GB

uv tool upgrade llm-wiki-kit updates it. The latest code, before a release: uv tool install git+https://github.com/GriffinHackley/LLM-WIKI.

An install from before 0.3.0 also provides wiki and has another name: wiki-cli, or llm-wiki-cli if it came from git or a clone. uv tool list shows it; remove it first with uv tool uninstall <name>. Wikis, their config and downloaded models are untouched.

Quick start

A new research wiki:

wiki new my-wiki --agent claude     # optional slash commands (also copilot, opencode); AGENTS.md works for any agent
cd my-wiki
# put a source in raw/, then ask your agent: "ingest raw/<file>"

A wiki about a codebase, in its own repo beside the code, or in a folder of the code's repo:

wiki new my-app-wiki --preset code --code ../my-app
# or, from the code repo: wiki new docs/wiki --preset code --code .
# then connect the code repo for its agents: wiki connect ../my-app --agent claude-code

A wiki for your writing: you write ideas, scenes and chapters in originals/ in Obsidian, the agent never touches them, and keeps a wiki of the stories:

wiki new my-stories --preset writing --agent claude
# open it in Obsidian (Templater, Linter, Obsidian Git), write, then ask: "ingest the storm scene"

An existing folder of notes (an Obsidian vault, a docs/ folder, a wiki an agent already keeps):

cd my-notes
wiki new . --agent claude           # drafts a .wiki-cli.toml from your pages, adds the agent files
wiki index refresh                  # index and embed (about 2 minutes per 500 pages on CPU)
wiki search "how do I configure the exporter?"

It also works with no configuration at all: every *.md file under the folder is a page, raw/**/*.txt is searchable source text, and every link is a links-to relation whose reason is the sentence around it.

Contributing

CONTRIBUTING.md covers development setup, the tests, the docs and releasing. PLAN.md is the current plan, and FUTURE_IDEAS.md the deferred ideas.

License

MIT; see LICENSE.