LLM-compiled knowledge bases for
any AI agent with awesome outputs.

1. Install the plugin

claude plugin install wiki@llm-wiki

2. Create a topic wiki

Pick any topic you're curious about:

/wiki init nutrition

This creates a hub at ~/wiki/ and your first topic wiki at ~/wiki/topics/nutrition/.

3. Research it

/wiki:research "gut microbiome and mental health" --wiki nutrition
# or just say it naturally:
/wiki research gut microbiome and mental health

Five parallel agents search the web from different angles (academic, technical, applied, news, contrarian), ingest the best sources, and compile them into cross-referenced wiki articles. Takes 2-5 minutes.

4. Ask your wiki a question

/wiki:query "how does fiber affect mood?" --wiki nutrition
# or naturally:
/wiki how does fiber affect mood?

The wiki answers from its compiled articles with citations.

5. Audit before you trust an output

/wiki:audit --wiki nutrition
/wiki:audit --artifact output/report-gut-brain.md

Audit rechecks the wiki layer, traces the output's evidence chain, flags drift, and will do fresh research if the local corpus is not enough to answer the trust question.

What to do next:

• /wiki:research "topic" --deep — 8 agents instead of 5, adds historical and data angles
• /wiki:research "topic" --min-time 1h — keep researching in rounds for an hour
• /wiki:research "topic" --plan — decompose into parallel research paths
• /wiki:audit --project nutrition-playbook — verify outputs and upstream wiki state together
• /wiki add https://example.com/article — fuzzy router detects the URL and ingests it
• /wiki what do we know about CRISPR? — fuzzy router detects the question and queries
• /wiki:lint --fix — clean up any structural issues

A typical research session flows through four stages:

Stage 1: Ask a question or pick a topic

llm-wiki auto-detects whether you're asking a question or naming a topic. Use the direct command or the fuzzy router:

/wiki:research "What makes long-form articles go viral?"   # direct command
/wiki research quantum computing                            # fuzzy router — same result
/wiki:research --mode thesis "fiber reduces neuroinflammation via SCFAs"  # thesis → for/against evidence

Stage 2: Agents search in parallel

5 agents (8 with --deep, 10 with --retardmax) search simultaneously from different angles — 2-3 web searches each, full-content fetch, quality scoring (1-5). A credibility pass deduplicates before ingestion.

Stage 3: Sources are ingested and compiled

Top sources are saved to raw/ (immutable — never modified after ingestion). Then the compilation pass synthesizes them into wiki articles under wiki/concepts/, wiki/topics/, and wiki/references/ with cross-references, confidence scores, and bidirectional links.

Stage 4: Gap report and follow-up

After each round, you see what's covered, what's still missing, and suggested follow-ups. If 2+ gaps remain, you're offered to close them in parallel:

### Close gaps?
1. Dose-response curves for wavelength specificity
2. Long-term safety data for daily exposure
3. Device comparison (clinical vs consumer panels)

Enter numbers (e.g. 1,2,4), "all", or "skip":

Multi-round research: Add --min-time 2h to keep researching in rounds, each drilling into gaps the previous round found. Add --plan to decompose into parallel paths upfront.

The hub (~/wiki/) is just a registry. No content — only wikis.json, _index.md, and log.md. All content lives in topic sub-wikis.

Topic wikis (~/wiki/topics/<name>/) are isolated research areas. Each has its own sources, articles, outputs, and Obsidian vault config. Isolation means researching quantum computing can't pollute your nutrition wiki.

Raw sources (raw/) are immutable. Once a paper, article, or data file is ingested, it's never modified. This is the audit trail — every claim in every article traces back to a source.

Wiki articles (wiki/) are LLM-compiled syntheses organized into three categories:

• Concepts — foundational ideas, mechanisms, theories
• Topics — specific subjects, comparisons, state-of-the-field
• References — tools, frameworks, data tables, lookup resources

Archive (topics/.archive/) is for whole topic wikis the user no longer wants in normal context. It preserves source history, articles, outputs, and logs while keeping old interests quiet by default.

Inventory (inventory/) is for durable operational state: actual items, source candidates, corpora, entities, open questions, tasks, watch items, and next actions. It is intentionally not evidence for factual claims.

Dataset manifests (datasets/) let the wiki index large or external data without copying it into raw/. Manifests can point to local paths, URLs, archives, samples, profiles, and query recipes.

Articles use dual-link format: [[wikilink]] for Obsidian + standard markdown links for everything else. Confidence scores (high/medium/low) reflect source quality and corroboration.

Indexes (_index.md) exist in every directory. They're derived caches — rebuilt automatically from file frontmatter. The agent reads indexes first and never scans blindly.

Outputs (output/) are generated artifacts: reports, slide outlines, study guides, implementation plans. They're built from wiki articles, so every output compounds on all prior research.

Audit walks that full artifact graph. It can trace an output back through the wiki state and raw sources it depended on, then escalate into fresh research when the stored evidence is stale or incomplete.

Use iCloud Drive when you want the same hub visible from multiple Macs. The important rule is that the shared config should describe the logical path, not one Mac's absolute user directory.

Set the hub with a portable path:

/wiki config hub-path "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"

This writes ~/.config/llm-wiki/config.json. New configs should keep hub_path and avoid writing a machine-specific resolved_path. Shared wikis.json entries should store topic paths such as topics/bitcoin, not /Users/alice/.../topics/bitcoin.

Each Mac needs its own app permission. iCloud sync can be healthy while macOS still blocks the agent process. Full Disk Access is local to the machine and to the exact launcher app.

• Claude Code from Terminal: grant Terminal.app access, then fully restart Terminal.
• Codex from VS Code, Cursor, iTerm, or another launcher: grant that launcher, not just Terminal.
• Codex over SSH: grant access to /usr/libexec/sshd-session, /usr/libexec/sshd-keygen-wrapper, and optionally /usr/sbin/sshd, then start a new SSH/Codex session.
• OpenCode or sandbox wrappers: grant the app permission and also allow the hub path in the tool's own sandbox config.

Fast diagnostic: run this from the same terminal, SSH session, or app that launches the agent:

python3 - <<'PY'
from pathlib import Path
hub = Path.home() / "Library/Mobile Documents/com~apple~CloudDocs/wiki"
for path in [hub, hub / "wikis.json", hub / "topics"]:
    try:
        print("STAT OK:", path)
        path.stat()
    except Exception as exc:
        print("STAT FAIL:", type(exc).__name__, exc)
    try:
        if path.is_dir():
            print("LIST OK:", [p.name for p in path.iterdir()][:10])
        else:
            print("READ OK:", path.read_text(encoding="utf-8")[:200])
    except Exception as exc:
        print("READ/LIST FAIL:", type(exc).__name__, getattr(exc, "errno", None), exc)
PY

How to read the result: STAT OK followed by READ/LIST FAIL ... Operation not permitted means the path is correct and macOS privacy is denying content access. Do not switch to ~/wiki or a machine-local resolved_path; grant the launcher permission and restart the session.

For broad topics, --plan decomposes your research into independent paths and runs them all in parallel:

/wiki:research "red light therapy" --plan --wiki redlight-therapy
# or naturally:
/wiki research red light therapy --plan

The agent generates a research plan and asks for confirmation:

## Research Plan — red light therapy

### Paths (will run in parallel)

1. **Mechanisms** — cytochrome c oxidase, wavelength specificity, dose-response
2. **Clinical evidence** — RCTs for skin, joints, wounds, cognition
3. **Devices** — LED vs laser, panel sizing, FDA clearances
4. **Criticisms** — placebo effects, publication bias, safety concerns

### Estimated: 4 paths x 5 agents = 20 parallel agents

Proceed? (y/n/edit)

On confirmation, all paths launch simultaneously. Each path runs its own 5-agent swarm. Sources are ingested in parallel (each path writes unique files), then a single compilation pass sees all sources at once for cross-path synthesis.

After the first round, a gap report shows what's still missing. Pick which gaps to close — they launch as another parallel batch:

### Close gaps?
1. Dose-response curves for wavelength specificity
2. Long-term safety data for daily exposure
3. Device comparison (clinical vs consumer)

Enter numbers (e.g. 1,2,4), "all", or "skip":

Combine with other flags:

• --plan --deep — 8 agents per path instead of 5
• --plan --min-time 2h — multiple plan-dispatch-compile cycles over 2 hours
• --plan --new-topic — create the wiki and research in one shot

Once your wiki has compiled articles, generate deliverables from them:

/wiki:output report --topic gut-brain --wiki nutrition
# or naturally:
/wiki write a report on gut-brain axis

Output types:

• summary — concise overview of a topic or the entire wiki
• report — detailed analysis with citations and evidence
• study-guide — structured learning material with key concepts and review questions
• slides — slide deck outline with speaker notes
• timeline — chronological view of events and developments
• glossary — term definitions extracted from articles
• comparison — side-by-side analysis of two or more subjects

Outputs are saved to output/ inside the topic wiki and indexed automatically. Every output builds on all compiled articles, so the more you research, the stronger every output gets.

Retardmax mode works here too — /wiki:output report --retardmax ships a rough draft immediately. Iterate later.

Cross-wiki context: Use --with to pull knowledge from another wiki into your output:

/wiki:output report --wiki nutrition --with article-writing
# Uses nutrition content + article-writing craft knowledge

Projects: Group related outputs into project folders with goals:

/wiki:project new rebuild-blog "Rewrite the company blog using wiki research"
/wiki:output report --topic gut-brain --project rebuild-blog

The output lands in output/projects/rebuild-blog/ with a WHY.md that captures the goal. Future outputs with the same --project flag accumulate there.

Research sessions fetch many URLs. By default, Claude Code asks for approval on each one. To skip these prompts, add WebFetch and WebSearch to your project's allow list:

In .claude/settings.local.json:

{
  "permissions": {
    "allow": [
      "WebFetch",
      "WebSearch"
    ]
  }
}

This pre-approves all web fetches and searches. You can also allow specific domains only: WebFetch(domain:arxiv.org).

If you prefer per-session approval, choose "Always allow" when Claude asks about the first URL — that covers the rest of the session.

LLM Wiki is a set of commands and a knowledge model that turns any LLM agent into a research engine and an append-only, Markdown-native wiki. It runs parallel multi-agent research, ingests URLs and files, archives old topic wikis that should stay preserved but quiet, tracks durable inventory state, indexes large datasets with manifests, compiles raw sources into synthesized articles with cross-references, audits whether those articles and outputs are still trustworthy, answers questions against the compiled knowledge, and generates artifacts like reports, slides, study guides, and implementation plans.

Inspired by Andrej Karpathy's LLM wiki concept.

Five install modes: Claude Code (native plugin via the llm-wiki marketplace), OpenAI Codex (marketplace plugin via codex plugin marketplace add nvk/llm-wiki, invoked with @wiki), OpenCode (instruction file via opencode.json), Pi (instruction file — best for local models with its minimal system prompt), and any other LLM agent via the portable AGENTS.md file.

The behavioral logic lives in a single wiki-manager skill shared across runtimes — Codex, OpenCode, and Pi trees symlink into the Claude source of truth so there is no fork. Drift is caught by self-healing sync tests.

By default in ~/wiki/ on your machine. You can relocate the hub to iCloud Drive, Dropbox, or any custom path:

/wiki config hub-path ~/Library/Mobile\ Documents/com~apple~CloudDocs/wiki

Config lives at ~/.config/llm-wiki/config.json. Shared hubs should use portable hub_path values and relative wikis.json topic paths; old machine-specific resolved_path values are treated only as fallback caches.

macOS privacy permissions are local to each Mac and to the exact process launching the agent. One machine can read the iCloud hub while another machine with the same Apple ID and same files returns Operation not permitted.

If stat succeeds for the hub but reading wikis.json or listing topics/ fails, the configured hub_path is correct. Grant Full Disk Access or iCloud Drive access to the launcher app, then fully restart that app and the agent.

For SSH-launched Codex sessions, Terminal.app permission is not enough because Codex is a child of sshd-session. Grant access to /usr/libexec/sshd-session, /usr/libexec/sshd-keygen-wrapper, and optionally /usr/sbin/sshd, then start a new SSH session.

Yes. Each topic wiki ships with its own .obsidian/ vault config and can be opened as an independent vault:

open ~/wiki/topics/nutrition/

Cross-references use a dual-link format:

[[gut-brain-axis|Gut-Brain Axis]] ([Gut-Brain Axis](../concepts/gut-brain-axis.md))

Obsidian reads the wikilink for graph view and backlinks; everything else (Claude Code, GitHub, plain text editors) follows the standard markdown link.

A research mode inspired by Elisha Long's retardmaxxing philosophy — act first, think later. Ten parallel agents, skip planning, cast the widest net, ingest aggressively, compile fast, lint later.

Available on /wiki:research --retardmax and /wiki:output --retardmax for when you want results now and will clean up afterward.

/wiki:research --mode thesis "<claim>" starts from a specific claim and uses it as a filter. Agents are split across supporting, opposing, mechanistic, meta/review, and adjacent — balanced by design.

Sources that don't relate to the claim's variables are skipped, which keeps the wiki tight. Output is a verdict: supported, partially supported, contradicted, insufficient evidence, or mixed.

With --min-time, round two focuses harder on the weaker side of the evidence — counter-weight against confirmation bias.

Yes. LLM Wiki is MIT-licensed. Source and releases at github.com/nvk/llm-wiki.

Compiling, querying, linting, and generating artifacts from an existing wiki work offline — everything is plain Markdown on your disk. Research and ingestion need internet since they fetch URLs and search the web.

Zero runtime dependencies. LLM Wiki uses only the built-in tools of the host agent (file read/write, web fetch, web search). The plugin itself is Markdown: command definitions, skills, and reference docs.

Optional: ask-grok-mcp for best-in-class tweet ingestion, tobi/qmd for local search beyond ~100 articles.

Claude Code: claude plugin update wiki@llm-wiki and restart.

Codex: codex plugin marketplace upgrade llm-wiki (or re-run the bootstrap helper for a local checkout).

AGENTS.md: re-curl from the master branch to replace your file.

See Install for the manual sync fallback when the plugin updater hits stale cache.