# LLM Wiki

> LLM-compiled knowledge bases for any AI agent. Export comprehensive Project Knowledge Checkpoints; apply personal specialist review methods without publishing them; run explicitly trusted adapters while their real inputs and runtime data stay external; capture and shape Ideas; and promote approved snapshots into Projects. Also includes compact queries, parallel research, sessions, ingestion, compilation, inventory, datasets, audits, and artifact generation. Ships for Claude Code, Codex, OpenCode, Pi, DS4, and portable agents. Obsidian-compatible. MIT-licensed.

Author: nvk · X: https://x.com/LLMWIKI · Source: https://github.com/nvk/llm-wiki · License: MIT

## What it is

LLM Wiki is a set of slash commands and a knowledge model that turns any LLM coding agent into a research engine and an append-only, Markdown-native wiki. Its Project Knowledge Checkpoints Export synthesizes related compiled knowledge across topics into a comprehensive fixed five-file repository handoff with exact coverage accounting, explicit omissions, and dry-run-first writes; semantic minimization and deterministic privacy sealing protect the export. Its optional personal specialist framework applies user-owned, instruction-only `SKILL.md` evidence and review methods stored under the locally resolved wiki hub through explicit per-topic allowlists and records version/hash provenance. Its private-adapter protocol invokes explicitly trusted local research tools through a machine-local registry, verifies their returned artifacts, and never imports results automatically. Its fuzzy Concept → Idea → Project workflow captures proposals under `inventory/ideas/`, researches and shapes them without implying commitment, then requires approval before freezing `BRIEF.md` in a linked Project. It also runs parallel multi-agent research (5-10 agents from different angles), collects provenance-rich catalogs, captures redacted session context and feedback, ingests and compiles sources, archives quiet topics, tracks inventory, indexes external datasets, answers questions, and generates artifacts.

It is inspired by Andrej Karpathy's "LLM wiki" concept (https://x.com/karpathy/status/2039805659525644595).

## Install

- Claude Code (native plugin): `claude plugin install wiki@llm-wiki` — best for full agentic research with 200K context.
- OpenAI Codex (marketplace plugin): run `codex plugin marketplace add nvk/llm-wiki`, then `codex plugin add wiki@llm-wiki`. Use `@wiki` for full workflows; in CLI/IDE type `$` or open `/skills` and select `wiki-query` for compact read-only lookups. For a local checkout: `./scripts/bootstrap-codex-plugin.sh --scope user --verify`.
- OpenCode (instruction file): load `plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md` for full workflows or `plugins/llm-wiki-opencode/skills/wiki-query/SKILL.md` for best-effort read-only queries. Web search requires `OPENCODE_ENABLE_EXA=1`.
- Pi (read-only query): `./scripts/pi-wiki-query`; use `./scripts/pi-ds4-wiki-query` for isolated local DS4. Both expose only `read,grep,find,ls` and disable unrelated discovery.
- Pi (full workflow): `pi --skill path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md`.
- Any LLM agent (portable query): copy `profiles/query-lite/SKILL.md` into the agent context or project as `AGENTS.md`. Use the root `AGENTS.md` only for complete write-capable research and maintenance.

## Chrome collaboration (development build)

- LLM Wiki for Chrome is a separate public-source extension and local Native Messaging companion at https://github.com/nvk/llm-wiki-chrome. Runtime grants, page data, credentials, result slots, and sockets remain local.
- Install the current development-only Homebrew build with `brew install --HEAD nvk/tap/llm-wiki-chrome`, then run `llm-wiki-chrome install`.
- Run `llm-wiki-chrome extension-path`, open `chrome://extensions`, enable Developer mode, choose Load unpacked, and select the printed directory. Pin LLM Wiki for Chrome.
- Register its fixed MCP server in Codex with `codex mcp add llm-wiki-browser -- llm-wiki-chrome mcp-server`, then start a new Codex session.
- Run `llm-wiki-chrome doctor` to verify the package and Native Messaging registration.
- To collaborate, open an HTTPS page and click the pinned extension icon. The exact tab receives a green outline and controlled marker; the side panel reports that the local agent was notified. Stop, close the tab, or navigate away from its origin to revoke the grant.
- The extension has no arbitrary JavaScript or raw-CDP tool, no persistent host permissions or `<all_urls>`, no ambient tab enumeration, and no repository storage of page content. Protected mutations still require local confirmation; consequential provider workflows should use targeted adapters with independent verification.
- Full installation, usage, security, and troubleshooting guide: https://llm-wiki.net/chrome/

## Published adapters

- Directory: https://llm-wiki.net/adapters/
- Google Docs Editing: https://github.com/nvk/llm-wiki-adapter-google-docs-editing
- Google Docs Editing v0.8.9: https://github.com/nvk/llm-wiki-adapter-google-docs-editing/releases/tag/v0.8.9
- The adapter plans exact replacements or one bounded append, applies tracked suggestions in Google Docs, and verifies the browser-visible result.
- It uses explicitly shared browser tabs through `llm-wiki-chrome` v0.1.0 or later. It does not use Google OAuth or the Google Docs API.
- Skill Factory: https://github.com/nvk/llm-wiki-adapter-skill-factory
- Skill Factory v0.1.0: https://github.com/nvk/llm-wiki-adapter-skill-factory/releases/tag/v0.1.0
- Invoke it explicitly with `wiki skill-factory <skill request>`. It has zero ambient routes and emits external disabled candidates plus source and pending evaluation records.

## Core commands

- `/wiki <natural language>` — fuzzy intent router (URLs → ingest, questions → query, etc.)
- `/wiki init <name>` — create a topic wiki at `~/wiki/topics/<name>/`
- `/wiki:idea new|list|show|develop|shape|promote|archive` — capture and catalog rough Ideas, research and challenge them, then explicitly promote an approved frozen brief into a linked Project
- `/wiki:portfolio [filters]` — show a live, read-only cross-topic view of canonical Ideas and active direct/promoted Projects without a catch-all store
- `/wiki:research <topic>` — 5 parallel agents; `--plan` (multi-path), `--deep` (8), `--retardmax` (10), `--new-topic`, `--min-time 1h`; repeat `--specialist <name>` at most three times or use `--no-specialists` for the baseline
- `/wiki:research --mode thesis "<claim>"` — thesis-driven: evidence for + against → verdict
- `/wiki:ingest <url|file|text>` — add sources; `--inbox` processes the inbox
- `/wiki:ingest-collection <source>` — bulk-ingest Git doc repos, BIP proposal sets, MediaWiki dumps/API sites, message archives, or Wayback CDX snapshots
- `/wiki:adapter add <local-path>` — register an existing explicitly trusted adapter with scoped read/write roots; llm-wiki never clones, updates, or publishes it
- `/wiki:adapter list|show|doctor` — inspect registrations and verify manifest/executable handshakes
- `/wiki:adapter route --intent <effect> --resource <url>` — resolve one healthy manifest-declared adapter and its private workflow guide without echoing the resource; ambiguity and drift fail closed
- `/wiki:adapter run <id> --request <json>` — run a scoped `llm-wiki-adapter/v1` request and verify artifact paths and hashes without automatic import
- `/wiki:adapter remove <id> --yes` — remove only the machine-local registration
- `/wiki:specialist init|create|refresh|list|show|validate` — manage personal, instruction-only specialist packages under the local hub at `HUB/.skills/`
- `/wiki:specialist enable|disable <name> --wiki <topic>` — change the explicit per-active-topic allowlist; there is no global default
- `/wiki:specialist suggest [--wiki <topic|all>]` — rank reusable bounded methods from current wiki indexes without creating or publishing them
- `/wiki:specialist apply <name> <question> --wiki <topic>` — apply one enabled, validated specialist and report its version/hash
- `/wiki:checkpoint create [slug] --repo <path> --audience private|team|public [--apply]` — preview, then optionally stage, seal, verify, and write a comprehensive five-file project handoff under `docs/knowledge/<slug>/`; writing does not authorize commit or publication
- `/wiki:checkpoint refresh <bundle> [--apply]` — verify first, re-run stored scope, and review source/claim/privacy changes before replacement
- `/wiki:checkpoint verify <bundle>` — read-only verification of bundle structure, hashes, comprehensive coverage, privacy findings, and exact overrides
- `/wiki:checkpoint import <bundle|repo> --wiki <name>` — verify and ingest one pinned checkpoint as bounded source evidence, never canonical truth
- `/wiki:collect "<things>"` — find, deduplicate, and catalog artifacts, examples, resources, media, memes, tools, entities, or source candidates
- `/wiki:collect "<things>" --scale tiny|small|medium|large|huge` — control write behavior by operational scale, not just row count
- `/wiki:collect "<things>" --media archive|thumbnail|reference` — download bounded originals by default for media-bearing collections; use thumbnails for previews or reference to opt out
- `/wiki:collect "<things>" --inventory corpus|records` — create one corpus record for large/media-heavy sets or per-item records for small durable sets
- `/wiki:inventory list|add|show|save-view|scan-outputs|migrate-output` — track durable items, source candidates, corpora, entities, watch items, and next actions
- `/wiki:dataset list|add|show|profile|sample|scan-outputs|migrate-output` — index large or external data with manifests, samples, profiles, and query recipes without copying the data
- `/wiki:archive list|topic|restore|peek` — archive whole topic wikis under `topics/.archive/`, restore them later, or search archived indexes without loading archived articles
- `/wiki:session status|disable|enable|capture|list|show|rehydrate|promote` — manage default-on redacted session capture, opt out, rehydrate context, or promote digests into topic raw notes
- `/wiki:feedback list|show|capture|promote` — review redacted user-feedback candidates, manually capture feedback, or promote durable lessons into topic raw notes
- `/wiki:compile` — raw sources → synthesized articles with cross-refs
- `/wiki:query <question>` — compact read-only, index-first Q&A with exact file citations; `--quick`, `--deep`, `--list`, `--resume`, `--include-archived`
- `/wiki:audit` — umbrella trust audit across wiki, outputs, provenance, and fresh research
- `/wiki:librarian` — focused wiki maintenance: score articles for staleness/quality and optionally propose topic-guide conventions
- `/wiki:ll` — extract lessons learned from current session into wiki; `--dry-run`, `--rules`
- `/wiki:plan <goal>` — wiki-grounded implementation plan; `--format rfc|adr|spec`
- `/wiki:output <type>` — summary, report, study-guide, slides, timeline, glossary, comparison
- `/wiki:assess <path>` — repo vs wiki vs market gap analysis
- `/wiki:librarian` — score articles for staleness and quality; `--passes conventions` writes proposal outputs without changing `schema.md`
- `/wiki:lint [--fix|--deep]` — structural and factual health checks; `--fix` can add a missing advisory topic guide
- `llm-wiki lint [path] --fail-on critical|warning|suggestion` — deterministic lint for automation with a configurable lowest severity that exits non-zero
- `llm-wiki schema status|adopt <wiki-root>` — local helper to inspect or add the human-owned `schema.md` topic guide
- `/wiki:retract [<source-path>]` — dry-run-first removal of a source and its downstream wiki references, including selected archives and sessions
- `llm-wiki retract --everywhere [--apply] [--remove-from-logs]` — clean up an accidentally saved sensitive value with hidden/stdin input, common encoded-form matching, explicit application, and local verification; `--remove-from-logs` removes complete matching `log.md` entries when their surrounding context is sensitive, while `--delete-binary-matches` opts into deleting whole matched binary files
- `/wiki:project` — create and manage committed delivery folders directly or from promoted Ideas; Projects own changing plans, artifacts, and outcomes

## Sessions and feedback

- Default-on, redacted operational memory lives under `HUB/.sessions/` or `.wiki/.sessions/` for local project wikis.
- Trusted hooks store harness metadata, current directory, git context, small redacted hook events, per-session state JSON, Markdown digests, and feedback candidates; full transcripts are not stored by default.
- A session ID such as `codex:abc123` is `<harness>:<native-session-id>`, a local lookup key that joins events, state, and a digest; it is not a secret, transcript, wiki, or topic ID.
- `session status|list|show|capture` inspects and creates digests; `session disable` opts out by writing `enabled: false`; `session enable` turns capture back on.
- `session rehydrate --cwd "$PWD"`, `--session-id`, or `--topic` returns compact context for the next agent turn without bulk-pasting a transcript.
- `session promote` and `feedback promote` are the boundary between operational memory and topic knowledge: selected digests or feedback candidates become distilled notes under topic `raw/notes/`.
- `feedback list|show|capture|promote` handles corrections, preferences, approvals, and plan acceptance; generic acknowledgements such as `ok`, `thanks`, and `cool` are ignored.
- Harness-session capture is separate from topic-level research crash state (`.research-session.json`, `.thesis-session.json`) and durable workflow provenance (`.session-events.jsonl`, `.session-checkpoint.json`).

## Trust model

`/wiki:librarian` keeps the `wiki/` layer in check. `/wiki:audit` is broader: it can trace outputs through `wiki/` and `raw/`, detect drift, inspect provenance, and do new research when the stored corpus is not enough to decide whether something is trustworthy.

## Architecture (one-liner)

Hub at `~/wiki/` is a lightweight registry (`wikis.json`, `_index.md`, `log.md`, optional personal `.skills/`, and optional `.sessions/` including `feedback/`). All active content lives in isolated topic sub-wikis at `~/wiki/topics/<name>/`, each with `inbox/`, `inventory/ideas/`, `datasets/`, `raw/`, `wiki/{concepts,topics,references}/`, `output/projects/`, plus `schema.md` as a human-owned topic guide and its own `.obsidian/` vault config. Archived topic wikis live under `~/wiki/topics/.archive/<name>/` and are hidden by default.

Project Knowledge Checkpoints Export writes outside the hub to a target repository at `docs/knowledge/<slug>/`. Each bundle has exactly `index.md`, `project-knowledge.md`, `sources.md`, `checkpoint.json`, and `privacy-report.json`; raw notes, sessions, prompts, local paths, usernames, secrets, and private data-plane locators are not dump formats.

## Key design properties

- v0.25.0 portable multi-runtime reliability — configurable deterministic lint exit gates, Claude/Codex/OpenCode packaging without `rsync`, safer Windows session-state and path handling, `HOME`-aware hub discovery, portable benchmarks, and isolated Codex installs that respect the selected `CODEX_HOME`.
- v0.24.3 privacy-sensitive log retraction — `--remove-from-logs` reports and removes complete matching `log.md` entries while preserving unrelated history; without the flag, the default continues to replace only matching values with `[RETRACTED]`. Apply mode performs the usual verification scan.
- v0.24.0 Project Knowledge Checkpoints Export — comprehensive cross-topic repository handoffs with exact per-input word counts, source-to-section mapping, explicit omissions, and a 20% thin-output floor. Semantic privacy minimization, deterministic sealing and verification, and exact audience-bound override attestations protect the export. Create and refresh are dry-run-first; verify is read-only; import is bounded evidence; no checkpoint operation silently authorizes commit, publication, or promotion.
- v0.23.0 personal specialist framework — optional user-owned `SKILL.md` review methods under the locally resolved wiki hub, explicit per-topic allowlists, instruction-only validation, bounded selection during research, and name/version/SHA-256 provenance. The public release contains infrastructure only; personal packages, definitions, allowlists, and wiki-derived candidate reports are never bundled or published. Personal does not mean encrypted or strictly local: synced hubs remain synced, and loaded instructions enter the selected model context.
- v0.22.0 declarative private-adapter routing — registered adapters own provider-neutral intent and exact-URL routes plus their workflow guides. The public plugin resolves a unique healthy match before ingestion without echoing the resource; ambiguity, manifest drift, and missing guides fail closed, while provider-specific setup and recovery remain private.
- v0.20.0 governed remote writes — exact remote-resource allowlists, declared effects, approval bound to an immutable plan, expected revisions, stable idempotency keys, private verified receipts, and content-free terminal reports.
- Zero runtime dependencies — uses only the host agent's built-in tools.
- v0.19.0 private adapters — a mode-`0600` machine-local registry, portable `llm-wiki-adapter/v1` contract, manifest handshakes, sanitized environments, path scopes, timeouts, atomic receipts, and SHA-256 artifact verification. Adapter code and bulk outputs remain external; `wiki-safe` is a review hint, never an automatic import instruction.
- v0.18.0 hub-wide portfolio — read active Idea indexes and Project `WHY.md` files, render Ideas and Projects separately, recognize only explicit lineage, and keep Concepts out of the backlog until captured as Ideas.
- v0.17.1 accidental sensitive-data cleanup — dry-run-first local retraction across registered wikis, archives, and session context with hidden/stdin input, common encoded-form matching, explicit application, and verification.
- v0.17.0 Ideas workflow — catalog proposals per topic, derive maturity instead of synchronizing stage fields, and preserve an approved decision boundary before Project creation.
- Concept → Idea → Project keeps knowledge, incubation, and committed delivery distinct — an Idea preserves proposal lineage; explicit approval freezes `BRIEF.md`; the Project owns changing delivery truth.
- Fuzzy Idea prompts are first-class — “I have an idea,” “does it hold water,” “poke holes in it,” and “make the narrow version a project” route without requiring object names, paths, or slugs.
- Query Lite is a shared 2.8 KB read-only protocol — Claude `/wiki:query`, explicit Codex `$wiki-query`, Pi/DS4 launchers, OpenCode, and portable agents share bounded index-first retrieval and exact citations. Full workflows remain available for writes.
- v0.16.0 token-efficiency benchmarks — Claude query instructions are 72.54% smaller; query-only contexts are roughly 76-95% smaller than full workflows; controlled DS4 serialized provider payload fell 64.30%. Static budgets, AB/BA runners, and case/fixture SHA-256 gates reject regressions. OpenCode live behavior remains best effort.
- Dual-linking: `[[wikilink]] + [standard markdown link](path.md)` on every cross-reference — works in Obsidian, Claude Code, GitHub, and plain text editors.
- Raw is immutable during normal workflows; explicit user-directed retraction is the exception.
- Inventory is operational state, not factual evidence — it tracks items, candidates, corpora, watch records, and next actions.
- Collect outputs are discovery maps, not evidence — they preserve aliases, found-in-context provenance, media URLs, cached asset paths, hashes, source context, dedupe notes, and inventory recommendations before selected sources are promoted into raw/wiki/inventory/datasets.
- Dataset manifests index large/external data — samples, profiles, and query recipes live in the wiki; the underlying data stays where it belongs.
- Archive is quiet preservation — old topic wikis move to `topics/.archive/`, remain structurally maintainable, and stay out of default query/compile/research/collect/output context unless explicitly included.
- Topic guides are human-owned — `schema.md` records local vocabulary, relationship verbs, article subtypes, source boundaries, and adoption notes; `schema_state: advisory` means suggestions only.
- Session capture is operational memory — trusted hooks write redacted JSONL/state and Markdown digests under `.sessions/` by default; `session disable` opts out, `session rehydrate` restores compact context, and promotion into topic wikis is explicit.
- Feedback is candidate memory — high-signal corrections, preferences, approvals, and plan acceptance become redacted candidates under `.sessions/feedback/`; generic acknowledgements are ignored and promotion is explicit.
- Topic-isolated indexes — no cross-topic noise, but queries peek sibling wikis for overlap.
- Configurable hub — store on iCloud, Dropbox, or any custom path via `/wiki config hub-path`; shared hubs prefer portable `hub_path` config and relative `wikis.json` topic paths so they work across machines with different home directories.
- iCloud permission diagnostics — if macOS lets Codex stat the configured hub but denies reads or `topics/` listings, the CLI and skill guidance treat it as a launcher privacy-permission issue, not an invalid registry or reason to use a machine-local fallback path.
- v0.15.1 Codex 0.144 compatibility — non-interactive cached plugin installation, cache-aware verification, clean-home install regression tests, and corrected lessons-learned raw-note frontmatter.
- v0.15.0 topic guides — reframes `schema.md` as a human-owned topic guide, adds `llm-wiki schema adopt`, and updates librarian language to convention proposals instead of database migrations.
- v0.14.0 default topic guides — new wikis get advisory `schema.md` by default; older wikis remain valid and can adopt a starter safely.
- v0.12.0 feedback curator — redacted feedback candidates under `HUB/.sessions/feedback/`, generic acknowledgement filtering, review/list/show, manual capture, and explicit topic promotion.
- v0.11.1 session helper compatibility — automated session capture works on Python 3.9/macOS system Python as well as newer Python runtimes.
- v0.11.0 automated session capture — default-on redacted hook checkpoints under `HUB/.sessions/`, `session disable` opt-out, compact rehydration, and explicit topic promotion.
- v0.10.1 collector media downloads — `/wiki:collect` downloads bounded public binary media into `output/assets/collect-<slug>/` by default, records local paths and hashes in the catalog, and keeps `--media reference` as the no-download opt-out.
- v0.10.0 collector catalogs — `/wiki:collect` finds, deduplicates, and catalogs examples, artifacts, media, memes, tools, entities, and source candidates with aliases, found-in-context provenance, scale handling, and media-safe defaults.
- v0.9.0 topic archive lifecycle — archive, restore, and peek archived topic wikis while preserving raw sources, compiled articles, outputs, logs, and registry history.
- Append-only activity log at `log.md` in every wiki.

## iCloud shared hub notes

- Set the hub with `/wiki config hub-path "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"` so each Mac expands `~` locally.
- Keep shared `wikis.json` topic paths relative, for example `topics/bitcoin`, rather than absolute `/Users/<name>/...` paths.
- macOS privacy permissions are per-machine and per-launcher. Grant Full Disk Access or iCloud Drive access to the exact app launching the agent, then restart it.
- If Codex is launched over SSH, Terminal.app permissions do not apply; grant access to `/usr/libexec/sshd-session`, `/usr/libexec/sshd-keygen-wrapper`, and optionally `/usr/sbin/sshd`.
- Diagnostic pattern: `stat` succeeds for the hub, `wikis.json`, and `topics/`, but reading/listing fails with `Operation not permitted`. That means the path is correct and access is blocked by macOS privacy controls.

## Docs

- Home: https://llm-wiki.net/
- Adapters: https://llm-wiki.net/adapters/
- Chrome collaboration: https://llm-wiki.net/chrome/
- Install: https://llm-wiki.net/#install
- Quick start: https://llm-wiki.net/#quickstart
- Sessions: https://llm-wiki.net/#sessions
- Commands reference: https://llm-wiki.net/#commands
- Guides: https://llm-wiki.net/#guides
- FAQ: https://llm-wiki.net/#faq
- README (canonical): https://github.com/nvk/llm-wiki#readme
- Releases: https://github.com/nvk/llm-wiki/releases
- AGENTS.md (portable): https://raw.githubusercontent.com/nvk/llm-wiki/master/AGENTS.md

## Related

- learntoprompt.org (https://learntoprompt.org) — sister site by the same author. Practical guides for AI coding agent setup (Claude Code, Codex, OpenCode, Pi, llama.cpp, ollama, ds4), sandboxing with bondage + nono, secret management with envchain-xtra, and prompting workflows. Includes a dedicated llm-wiki guide at https://learntoprompt.org/guides/llm-wiki.html and a portable DS4 wiki profile guide at https://learntoprompt.org/guides/ds4-wiki-profile.html.