llm-wiki

Personal knowledge wiki managed by Claude. You feed context; Claude writes.

Markdown, git-backed, accessible from any Claude Code session via MCP.

Install

Run these one at a time, in order — each depends on the previous one finishing (the plugin must be installed before /reload-plugins picks it up, and /reload-plugins must complete before /llm-wiki:wiki-setup exists as a command). Don't paste them as a block.

/plugin marketplace add git@github.com:ts-eva/llm-wiki.git

/plugin install llm-wiki@llm-wiki-marketplace

/reload-plugins

/llm-wiki:wiki-setup

/llm-wiki:wiki-setup runs an interactive wizard: creates your wiki repo (defaults to current directory), asks for name/focus/remote/date format, initializes git, and automatically registers the MCP server (user-scoped, pointing at the wiki path you chose) so wiki tools are available in every Claude Code session.

Reconnecting an existing wiki (after reinstalling the plugin, or the MCP server just stopped working): run /llm-wiki:wiki-reconnect — it re-registers the server for your existing wiki without touching its content or re-running the whole setup wizard.

Manual MCP registration (fallback — only needed if the plugin isn't installed/loading at all, so /llm-wiki:wiki-reconnect isn't available either):

PLUGIN_DIR=$(grep -l '"name": "llm-wiki"' "$HOME"/.claude/plugins/marketplaces/*/.claude-plugin/plugin.json | head -1 | sed 's#/.claude-plugin/plugin.json$##')
claude mcp add llm-wiki --scope user --env WIKI_PATH="<absolute-path-to-your-wiki>" -- node "$PLUGIN_DIR/server/start.mjs"

Register the marketplace clone, not ~/.claude/plugins/cache/.../llm-wiki/<version>/: each plugin update creates a new versioned cache folder, so a registration there silently stays on the old server. server/start.mjs installs the server's npm dependencies on first run and again whenever package-lock.json changes, so there's no manual npm install step. It uses npm install --no-save (never rewrites the lockfile, so the marketplace clone stays clean for updates; never wipes node_modules, so an offline update keeps the old working deps) under a lock, so several sessions starting at once don't race. --scope user matters: without it the server only registers for the current project.

Requirement: GitHub SSH key with access to ts-eva/llm-wiki.

Updating: after /plugin update llm-wiki@llm-wiki-marketplace, run /reload-plugins — the SessionStart hook is declared in .claude-plugin/plugin.json and only takes effect once the plugin reloads (or in the next new session). The MCP server reports the plugin version in its serverInfo, so you can tell which build a session is running.

Development: cd server && npm test (node's built-in runner, no dev deps). Try a working copy end to end without installing it, against a throwaway wiki:

claude -p "/llm-wiki:wiki-process auto" --plugin-dir . --strict-mcp-config \
  --mcp-config '{"mcpServers":{"llm-wiki":{"command":"node","args":["'"$PWD"'/server/start.mjs"],"env":{"WIKI_PATH":"/tmp/test-wiki"}}}}' \
  --permission-mode acceptEdits --allowedTools "mcp__llm-wiki" Read Write Edit Agent "Bash(git:*)"

Daily workflow

Drop files into sources/                     ← zero tokens, any time (any name — they get normalized)
/llm-wiki:wiki-add interesting article url   ← save a URL or paste text (searchable at once)
/llm-wiki:wiki-session                       ← capture today's Claude session

daily job (/llm-wiki:wiki-schedule)          ← processes new notes; free on idle days
/llm-wiki:wiki-process                       ← same, on demand: Haiku tags, Sonnet organizes
/llm-wiki:wiki-ask how does X work?          ← synthesized answer from your wiki
/llm-wiki:wiki-digest this week              ← what did I learn?

Commands

Command	Model	What it does
/llm-wiki:wiki-setup	free	First-time setup wizard
/llm-wiki:wiki-reconnect [path]	free	Re-register the MCP server for an existing wiki (e.g. after reinstalling the plugin)
/llm-wiki:wiki-add [text|url]	free	Save content to sources/ via save_source (searchable at once)
/llm-wiki:wiki-process [auto]	Haiku + Sonnet	Batch prepare + tag + organize everything new, changed or removed in sources/; auto = unattended
/llm-wiki:wiki-schedule [install HH:MM|uninstall|status]	free	Daily processing job (crontab)
/llm-wiki:wiki-session [topic]	Sonnet	Summarize current Claude session → sources/
/llm-wiki:wiki-ask [question]	session model	Search, read the top hits, answer
/llm-wiki:wiki-search [query]	free	search_wiki: ranked pages and sources
/llm-wiki:wiki-digest [range]	Haiku	"What did I learn this week/month?" from get_recent
/llm-wiki:wiki-stats	free	wiki_stats dashboard: pages, tags, pipeline status
/llm-wiki:wiki-retag	Haiku	Analyst proposes tag merges; rename_tag applies them
/llm-wiki:wiki-link	free	find_unlinked_mentions, then add the links you approve
/llm-wiki:wiki-autotag	Haiku	Tag only (no organizing); the daily job supersedes it
/llm-wiki:wiki-open	free	Open wiki vault in Obsidian
/llm-wiki:wiki-convert [format]	free	Switch between standard and Obsidian link format
/llm-wiki:wiki-commit	free	Manual git commit

MCP tools (available in every Claude session)

Tool	Description
search_wiki(query?, tags?, include_sensitive?, limit?)	Word search over pages and sources (unprocessed notes too). Every word must match, any order, any field; otherwise best partial matches, flagged. Ticket keys (HW-2538) stay whole. Tags filter every result
get_page(slug)	Fetch a page by slug
list_pages(type?, tag?, include_sensitive?)	List pages, filterable
list_tags()	Canonical tag list
save_source(content, title?, source_url?, type?, new?)	The write path into sources/ — /llm-wiki:wiki-add and /llm-wiki:wiki-session call it. Applies the naming rules, writes created:/updated:, commits; same title updates the note. A new title close to an existing note is refused with the similar titles (merge, or new: true)
get_recent(days? | since?, until?)	log.md entries plus sources created/updated in the window, processed or not
get_backlinks(source_file)	Pages referencing a source file
wiki_stats()	Dashboard numbers
find_unlinked_mentions()	Page titles mentioned in prose but not linked
rename_tag(from, to)	Rename/merge a tag in page frontmatter, index entries, tags.md
source_status()	Pipeline state of sources/ (untagged, changed, unorganized, removed, ignored) plus wiki, the path the server serves
prepare_sources()	Pipeline: backfill created:, rename untagged notes to the naming rules, relink [[wikilinks]]; returns files, the paths it changed
get_source_entries(files) / write_source_entry(file, entry) / set_wiki_pages(file, pages)	Pipeline: read/write single sources/index.md entries, so no agent loads the whole (large) file
commit_files(files, message)	Commit exactly these paths (a trailing / = every change under that folder), never git add -A. Built in a temporary index under a lock, so your staged files and other sessions' edits stay out, and case-only renames work on macOS
mark_sources(files, stage)	Pipeline: stamp entries organized / baseline / remove
append_log(entries) / remove_page(slug)	Pipeline: log.md append without reading it; delete an orphaned page and its index line

Tool results are compact JSON (no indentation) — they're model context.

Two ways to use

Ambient — MCP tools available in any Claude session. Drop a note, search, ask a question, save something interesting — without leaving your current project.

Direct — cd <wiki-path> && claude. CLAUDE.md loads automatically. Best for bulk ingestion, restructuring, or anything that needs the full wiki in context.

As Claude memory — automatic, via a SessionStart hook (hooks/session-start.js). Every session starts with the wiki's location, its most recent sources (newest by updated:/created: frontmatter, not file mtime, which pipeline steps churn), how many source changes await processing, and the naming rules already in context, so Claude recalls before it answers and names files correctly the first time. Instructions in ~/.claude/CLAUDE.md alone are not enough: they're prose the model can skip, and it does — a hook always fires.

The hook resolves the wiki the same way the MCP server does ($WIKI_PATH, else the path registered for llm-wiki in ~/.claude.json, else ~/wiki), prints nothing when no wiki is found, and runs in ~150 ms. Because the hook carries the rules, the ~/.claude/CLAUDE.md block /llm-wiki:wiki-setup writes is a two-line pointer — repeating the rules there would cost tokens every session for nothing. Session captures via /llm-wiki:wiki-session feed into the pipeline, which organises them into the right wiki pages — closing the loop.

Wiki structure

<wiki-path>/
├── CLAUDE.md            # Claude's operating manual
├── config.yaml          # Your settings
├── log.md               # Append-only change history
├── backlinks.md         # Source → page inverse index (standard mode)
├── .obsidian/           # Obsidian config (Obsidian mode only)
├── sources/             # Raw material — drop files here
│   └── index.md         # Pipeline-owned entry per source (tags, key points, hashes) — read/written only via tools
└── wiki/
    ├── index.md         # Master navigation index
    ├── tags.md          # Canonical tag list
    ├── schema.md        # Page type and frontmatter reference
    └── pages/           # Wiki pages

config.yaml

wiki:
  name: "My Wiki"
  focus: "general personal and work notes"
  author: "Your Name"
  link_format: standard       # standard | obsidian
  date_format: "MM/DD/YYYY"   # date display — YYYY-MM-DD for ISO, DD/MM/YYYY for European

git:
  auto_commit: true    # at session end, commit the wiki files that session edited
  auto_push: false     # push after commit (requires remote)
  auto_pull: false     # pull at session start (requires remote)

mcp:
  path: "/absolute/path/to/wiki"

Agents

Agent	Model	Role
wiki-tagger	Haiku	First pass: reads raw source, writes its entry with write_source_entry
wiki-curator	Sonnet	Fetches its entries with get_source_entries, writes pages, records them with set_wiki_pages / append_log / remove_page
wiki-analyst	Haiku	Judgment calls only: tag-merge proposals, digest narrative

Pipeline (/llm-wiki:wiki-process)

1. Prepare — prepare_sources stamps missing created: dates (earlier of git first-add and file birth time) and renames untagged notes that break the naming rules: uppercase, - separators, legacy session-…-2026-09-16 slugs (ticket keys like hw-2409 survive). Renames use git mv (case-only renames are otherwise invisible to git on macOS) and rewrite [[wikilinks]] to the old name. Tagged and ignore: true notes are never renamed.
2. Detect — source_status hashes every source against its entry: new, changed, unorganized, removed. Headers inside the index's <!-- --> format comment are ignored.
3. Tag — wiki-tagger (Haiku), batches of 25.
4. Organize — wiki-curator (Sonnet).
5. Commit once — commit_files with the sources this batch processed, sources/index.md, wiki/, log.md and backlinks.md; push only if git.auto_push. Never git add -A: other sessions' in-progress edits and anything you staged stay out of pipeline commits.

The wiki path comes from the server (wiki in source_status), never from the current directory or CLAUDE.md, so running it from any project folder is safe.

Daily processing (optional)

/llm-wiki:wiki-schedule install 16:30

Adds a crontab line (tagged # llm-wiki-process) running scripts/nightly-process.sh. Pick a time the machine is awake: cron skips runs while asleep.

• Free when idle: it counts pending work with plain node and exits without starting Claude when nothing is new.
• Otherwise it runs claude -p "/llm-wiki:wiki-process auto" in the wiki folder, with the MCP server pinned to that same wiki and a narrow tool allowlist (llm-wiki tools, Read/Write/Edit, Agent, git).
• Auth under cron: claude -p can't read the login keychain from cron and needs HOME/USER/LOGNAME/TMPDIR. The script sets those and reads a claude setup-token token from LLM_WIKI_TOKEN_FILE (default ~/.claude/.oauth-token, chmod 600). Set LLM_WIKI_TOKEN_FILE when running install to bake a different path into the cron line.
• Overlap: a mkdir lock skips a run if the previous one is still going.
• Log: ~/.claude/llm-wiki-process.log. Remove with /llm-wiki:wiki-schedule uninstall.

Session-end commit

With git.auto_commit: true (the default), the SessionEnd hook commits the wiki files the ending session edited, in the background so exit stays fast:

• Only its own files: the paths its Write/Edit calls (subagents included) pointed at inside the wiki, committed with commit_files. Edits another session has in progress, and anything you staged, are left alone, so several sessions can run at once.
• Nothing changed: no commit. save_source and /llm-wiki:wiki-process commit their own work, so they're already clean by then.
• Files changed through Bash aren't tracked; commit those by path.
• Pushes after the commit when git.auto_push is true and a remote exists.
• Log (commits and errors only): ~/.claude/llm-wiki-commit.log.

Obsidian support

Requires Obsidian installed. Run /llm-wiki:wiki-setup and choose Obsidian mode to get [[wikilinks]], graph view, and Dataview support. Switch any time with /llm-wiki:wiki-convert. Standard mode renders everywhere (GitHub, VS Code, Warp).

Obsidian does not need to be open for any wiki commands to work — all commands use Claude and file I/O directly. Open Obsidian when you want to browse the graph, run Dataview queries, or read notes visually.

In Obsidian mode the vault root is the wiki folder itself (not a subfolder), so sources/ files appear in the graph alongside wiki pages — you can see which source files generated which pages. Wiki pages link to sources via [[sources/filename]] wikilinks that Obsidian renders as graph edges.