llmwiki
New in 1.5 — Wiki behavior separated from storage.
1.5.0 is the stable release on npm latest: npm install --global llm-wiki-compiler.
- Same wiki, cleaner foundation. Compilation, retrieval and review now describe what they need from storage instead of touching files directly. Your existing wiki keeps its layout and works as before, with no migration.
- Converse gateways. Use a bearer-authenticated Converse endpoint as your provider.
- Pages that fit the source.
llmwiki compile --no-pages-limitdrops the fixed concept count. - A viewer that keeps up.
llmwiki viewrefreshes after a compile without a restart and reads its port fromLLMWIKI_VIEW_PORT.
New in 1.4 — Review answers before publishing.
Review generated answers and their citations before they become pages, approve a batch of pages in one operation, and trace an answer back to the pages that informed it. Release notes · Upgrade guide
New in 1.3 — A fresh look for your wiki.
Meet Scientific Clay, with soft surfaces and rounded typography, and Minimal, which follows your system’s light or dark setting. Switch instantly between four themes, including Nebula Light and Dark.
The same page in a demonstration wiki. Click either screenshot for a closer look.
Recursive source folders, path exclusions, project-specific compile instructions, and storage for larger embedding indexes.
New in 1.2 — Explore your records and their evidence.
Browse profile-defined categories, declared fields, connected records, provenance, and supporting source passages in the local viewer.
New in 1.1 — Publish and share domain templates.
Create signed template distributions, discover them through explicitly trusted catalogs, and install or update them with compatibility checks.
New in 1.0 — Configurable Lifecycle Profiles.
Build a knowledge system around the way you work. Define your records, relationships, review gates, and workflows in one validated profile. Start with AutoSci for research or Newsroom for editorial work, or create your own.
Explore Configurable Lifecycle Profiles →
What llmwiki does
Compile raw sources into an interlinked, citation-traceable markdown wiki that agents and humans can browse, query, lint, export, and reuse. The default profile preserves the classic concepts-and-queries layout; optional profiles add domain-specific types and workflows without adding domain branches to the compiler.
llmwiki implements the LLM Wiki pattern: instead of re-discovering knowledge from raw files at query time, compile it once into durable pages that accumulate structure, provenance, review state, and retrieval metadata over time.
When to use this repo
Use llmwiki when you need a persistent knowledge base from raw material:
- Compile papers, notes, READMEs, transcripts, PDFs, images, or web pages into typed wiki pages.
- Give agents a stable, citation-aware context pack instead of a pile of loose files.
- Keep generated knowledge auditable with source citations, review queues, freshness checks, and quality gates.
- Browse the result locally, query it from the CLI, expose it over MCP, or embed it through the SDK.
- Exchange compiled knowledge with other tools using Open Knowledge Format (OKF), JSON, JSON-LD, GraphML, Marp, and
llms.txt.
Do not use llmwiki as a general static-site generator, a heavy ontology database, or a replacement for ad-hoc search over fast-changing raw logs. It is strongest when source knowledge is worth compiling, reviewing, and reusing.
What you get
- Compiled wiki, not chunks. A two-phase LLM pipeline extracts concepts, then generates typed pages:
concept,entity,comparison, andoverview. - Configurable Lifecycle Profiles. A fail-closed
.llmwiki/profile.jsoncan declare entity schemas, typed relations, lifecycle state machines, transition requirements, workflows, artifacts, connectors, content tiers, and retrieval policy. - Installable domain templates.
llmwiki template init autoscicreates a research project with papers, ideas, experiments, manuscripts, evidence artifacts, workflows, and Crossref import.newsroomdemonstrates the same machinery for editorial work. - Runtime trust gates. Relation, evidence, artifact, and human/agent gates are enforced by the write path rather than left as prompt conventions; standing lint detects drift after the fact.
- Citation-traceable output. Paragraphs and claims cite source files and line ranges, and
llmwiki lintvalidates the links. - Hybrid retrieval. Semantic chunk search, BM25 reranking, and wikilink graph expansion build compact evidence packs for queries and agents.
- Local viewer.
llmwiki viewopens a read-only browser UI with search, page metadata, graph exploration, source-freshness badges, and citation chips. - Review policy. Generated pages can be auto-held for review when confidence, contradiction, schema, or provenance rules trip.
- Freshness repair.
llmwiki lintandllmwiki nextsurface stale/orphaned pages;llmwiki refresh --stalerepairs changed knowledge without compiling unrelated new sources. - Eval harness.
llmwiki evalreports health score, a per-page health distribution that flags the worst pages, wikilink-graph health, citation coverage/precision, corpus stats, regression deltas, and optional judge-model citation support. - MCP server.
llmwiki serveexposes ingest, compile, query, lint, read, status, eval, context-pack, and OKF exchange tools to MCP-compatible agents. - SDK.
createWiki({ root })drives ingest, compile, query, context, status, export, eval, and OKF import/export from TypeScript without shelling out. - Open Knowledge Format exchange. Export and import OKF bundles for portable, markdown-native knowledge exchange. External OKF imports are staged through the review queue by default; trusted bundles can be written live explicitly.
Configurable Lifecycle Profiles (CLP)
CLP turns llmwiki's knowledge compiler into a reusable substrate for domain-specific knowledge systems. A validated .llmwiki/profile.json is the single contract for:
- typed entities, fields, and directed relations;
- lifecycle states, transition evidence, and trust gates;
- multi-stage workflows and declared actions;
- hash-pinned artifacts and first-party connector bindings; and
- content tiers and retrieval behavior.
These rules are enforced by the runtime, not left as prompt conventions. The CLI, SDK, MCP server, viewer, context builder, lint, status, export, and OKF exchange surfaces all operate from the same profile contract. Invalid profiles and writes that bypass a declared gate fail closed.
CLP is backward-compatible by construction: a project without .llmwiki/profile.json uses the built-in default concepts-and-queries profile and preserves the pre-1.0 behavior. You can start three ways — scaffold your own profile, install a built-in or local template, or install a signed template from a trusted tap:
# author your own profile, one entity type at a time
llmwiki profile init research --entity paper
# or install a built-in or local declarative template
llmwiki template list
llmwiki template inspect autosci
llmwiki template init autosci
llmwiki profile validate
llmwiki workflow list
autosci is a practical research system with papers, ideas, experiments, manuscripts, evidence artifacts, workflows, and Crossref ingestion. newsroom applies the same generic machinery to articles, desks, bylines, and editorial workflows. Templates contain configuration and examples, never executable plugin code.
Templates can also be distributed securely. Publishers build signed, offline distributions with llmwiki template publish — Ed25519 signing, key rotation, and package revocation — and verify them with template publish verify. Consumers add explicitly trusted taps, discover and inspect signed catalogs, and install or update templates with continuity, revocation, and compatibility checks enforced under lock.
Read the CLP concept guide, follow the AutoSci research workflow, or explore the Newsroom editorial workflow.
Karpathy's LLM Wiki pattern
Andrej Karpathy described the LLM Wiki pattern as a way to turn raw material into compiled knowledge that future agents can reuse. llmwiki is a concrete compiler for that pattern.
The key shift is moving work from query time to compile time. Traditional RAG repeatedly retrieves raw chunks and asks the model to reconstruct relationships for each question. llmwiki first turns sources into typed, interlinked pages with citations, metadata, and review state. Queries, context packs, exports, and MCP tools then operate over that compiled artifact.
That makes llmwiki useful when knowledge should compound: concepts shared across sources become one page, saved answers become future context, stale pages can be detected and repaired, and agents can consume a stable evidence pack instead of re-reading the same raw files from scratch.
See docs/concepts/karpathy-pattern.mdx for the deeper explanation.
Agent decision guide
Use llmwiki for a reusable corpus of documents, research, or project knowledge.
Start with wiki_status and get_context_pack when you need evidence for your
own reasoning. A one-off file read or live web search does not need a wiki.
The npm package includes an Agent Skill. See the
MCP setup guide for installation.
The MCP Registry identifier is io.github.atomicstrata/llmwiki; the npm package
is llm-wiki-compiler and the executable is llmwiki.
Choose the entry point that matches the task:
Quick start
npm install -g llm-wiki-compiler
export ANTHROPIC_API_KEY=sk-...
# or choose another provider:
# export LLMWIKI_PROVIDER=openai
# export OPENAI_API_KEY=sk-...
llmwiki quickstart ./notes.md
llmwiki query "what are the key ideas?"
llmwiki view --open
quickstart ingests one source, compiles pages, and opens the viewer. Inside an existing project, run llmwiki next when you want the safest next action.
To start with a domain model instead of the default concepts-and-queries layout:
mkdir research-wiki && cd research-wiki
llmwiki template inspect autosci
llmwiki template init autosci
llmwiki profile validate
llmwiki workflow list
Template installation is for a new or empty typed project. It materializes the chosen profile into .llmwiki/profile.json; normal project loading never depends on a template registry or lockfile.
Demo
Try it on any article or document:
mkdir my-wiki && cd my-wiki
llmwiki quickstart https://en.wikipedia.org/wiki/Andrej_Karpathy
llmwiki query "What terms did Andrej coin?"
The examples/basic/ directory includes a small pre-generated wiki you can inspect without an API key.
Core commands
Full command docs live in docs/cli/.
Open Knowledge Format
llmwiki is an Open Knowledge Format (OKF) producer and consumer. OKF is a Google Cloud initiative for sharing compiled knowledge as portable markdown files with structured frontmatter.
llmwiki export --target okf --out ./dist/okf
llmwiki import --okf ./dist/okf --dry-run
llmwiki import --okf ./dist/okf
OKF import is intentionally review-first: untrusted bundles become review candidates, not live wiki pages. The importer preserves foreign OKF metadata, stores llmwiki provenance under x-llmwiki, and re-exports imported pages honestly after local edits, including safe original nested paths.
See docs/guides/open-knowledge-format.mdx, docs/cli/export.mdx, and docs/cli/import.mdx.
What llmwiki creates
A project has raw inputs in sources/, compiled markdown in wiki/, and compiler state under .llmwiki/:
sources/
raw source files
wiki/
concepts/ compiled pages
queries/ saved answers
<entity>/ profile-declared typed pages
graph/ typed relation and audit-event stores
outputs/ derived workflow projections
index.md generated TOC
.llmwiki/
profile.json active domain contract
template-lock.json advisory install provenance
config.json review policy and source selection
schema.json page-kind/cross-link policy
state.json source hashes and ownership
candidates/ held review candidates
workflows/ signed workflow run state
eval/ quality history and thresholds
artifacts/ hash-pinned profile-declared files and manifests
log.md activity journal
Compiled pages are plain markdown with YAML frontmatter, plus enough metadata for agents to reason about citations, freshness, confidence, contradictions, and review state. See docs/concepts/wiki-model.mdx.
Sources are top-level Markdown files by default. Opt into nested source folders and exclusions in project config. Excluding a compiled source retires its contribution on the next ordinary compile without deleting the source file.
Agent integration
MCP
Run:
llmwiki serve --root /path/to/wiki-project
MCP clients can ingest sources, compile, query, search pages, read pages, lint, run eval, inspect status, request context packs, and exchange OKF bundles. Read-only tools work without provider credentials; LLM-backed tools validate provider credentials at call time. The run_eval tool runs its fast suite without a provider; its full suite (which LLM-judges citation support) requires one.
See docs/guides/mcp-agent-integration.mdx.
SDK
import { createWiki } from "llm-wiki-compiler";
const wiki = createWiki({ root: "/path/to/wiki-project" });
await wiki.ingest({ source: "./notes.md" });
await wiki.compile();
const answer = await wiki.query({ question: "What changed?" });
See docs/guides/sdk.mdx. llm-wiki-compiler is the
supported entry point; its scoped supporting packages are implementation
dependencies. The local workflow engine and an application's own coordinator
are two tiers, not a migration path: llmwiki workflow runs a profile-declared
workflow through local invocations with persisted state between sessions.
An external coordinator calls the same SDK without using local workflow
execution methods; the supporting engine dependency remains installed. See
SDK package selection,
SDK upgrade notes, and
When to use the local engine.
Configuration
Minimum requirement: Node.js 24 or newer.
The default provider is Anthropic:
export ANTHROPIC_API_KEY=sk-...
Provider selection is environment-driven:
See docs/configuration/providers.mdx and docs/configuration/environment-variables.mdx.
Quality and safety model
llmwiki is designed for auditable generated knowledge:
- Review before write. Use
compile --reviewor.llmwiki/config.jsonreview policy to hold risky pages as candidates. - Profile floors are runtime checks. Field contracts, lifecycle transitions, relation counts, evidence, and artifact requirements are enforced across page, lifecycle, workflow, import, and approval write surfaces.
- External connector data is untrusted. First-party connectors use confined fetches and stage fenced review candidates; approval is pinned to the exact body the operator reviewed.
- Artifacts are content-addressed evidence. Artifact reads and writes are path-confined, size-capped, schema-checked, and verified against hash-pinned references.
- Fail-closed config. Invalid review-policy config aborts compile instead of silently disabling review.
- Source confinement. Source snippets and import/export paths are confined to the project.
- Freshness is explicit. Pages can be fresh, stale, orphaned, or unverified; stale pages are flagged and repairable. The JSON export is active-page-only: it carries freshness for live pages (
fresh/stale/unverified); computed-orphaned pages (all sources deleted) surface only as lint and viewer signals and are dropped from the export. - Imported compiled knowledge is staged by default. External bundles go through the review queue unless explicitly trusted.
- CI gates are supported.
llmwiki lintandllmwiki evalcan enforce quality thresholds.
See docs/configuration/review-policy.mdx, docs/troubleshooting/stale-pages.mdx, and docs/guides/ci-quality-gates.mdx.
Scale and what works
llmwiki is still early software, but it is no longer a toy pipeline for a handful of notes.
- Incremental compilation means unchanged sources do not flow back through the LLM.
- Parallel compile runs concept extraction and page generation concurrently under a configurable cap (
--concurrency/LLMWIKI_COMPILE_CONCURRENCY), cutting wall-clock on large compiles. - Chunk-level embeddings narrow large wikis before BM25 reranking and graph expansion.
- Content-hash-aware embedding updates avoid recomputing vectors for unchanged pages and chunks.
- Batch embedding sends page and chunk vectors to the provider in batches rather than one request at a time, cutting latency on cold starts and large refreshes.
- Binary embedding storage automatically handles stores above the 64 MiB JSON limit; smaller stores can opt in. Binary selection persists, and retrieval still loads the index into memory. See limits, recovery and downgrade guidance.
- Cached citation judgements make repeated
eval --suite fullruns cheaper. - Lexical fallback keeps query/context workflows usable when the active provider has no embedding endpoint.
- Prompt budgeting and ingest truncation metadata make large sources explicit instead of silently pretending they fit.
The current sweet spot is a durable project or domain wiki: research folders, codebase docs, team handbooks, standards, design notes, decision logs, or curated source packs. The less ideal fit is a high-churn firehose where raw search is enough and compiled structure would go stale faster than it can be reviewed.
Documentation
The full docs site source is in docs/:
- Start here:
docs/introduction.mdx - Quickstart:
docs/quickstart.mdx - Installation:
docs/installation.mdx - Karpathy's LLM Wiki pattern:
docs/concepts/karpathy-pattern.mdx - How the compiler works:
docs/concepts/how-it-works.mdx - Wiki model:
docs/concepts/wiki-model.mdx - Configurable Lifecycle Profiles:
docs/concepts/configurable-lifecycle-profiles.mdx - AutoSci research workflow:
docs/guides/autosci-research-workflow.mdx - Newsroom editorial workflow:
docs/guides/newsroom-editorial-workflow.mdx - Profile templates:
docs/configuration/profile-templates.mdx - CLI reference:
docs/cli/ - Open Knowledge Format:
docs/guides/open-knowledge-format.mdx - MCP integration:
docs/guides/mcp-agent-integration.mdx - SDK:
Preview the docs locally with Node 24:
cd docs
volta run --node 24 npx mint dev --port 3001
Current release
Stable release 1.5.0: The current release on the npm latest dist-tag. It ships the same code as the 1.5.0-rc.1 candidate.
- Wiki behavior is separated from local file I/O. The code that decides what a wiki does, such as which sources changed, what a page links to, or whether a review candidate can be approved, used to open, lock and rename files inline. Those rules now work through focused storage boundaries, and the current local-disk layout is their first implementation. Locking, atomic writes, path confinement, symlink handling and recovery stay in the local implementation. For you, nothing changes: the same directory layout, file formats, defaults, and CLI, SDK and MCP entry points, with no migration or storage setting. This is groundwork for alternative storage and for testing wiki logic independently of the disk; it does not add hosted storage, a remote backend, role-based permissions, or a public storage-plugin API. See
ARCHITECTURE.md. - Converse gateway provider.
LLMWIKI_PROVIDER=converseconnects to a bearer-authenticated endpoint speaking the Converse protocols, with streaming and structured extraction. Its AWS protocol packages are optional and load only when you select it. See provider setup. --no-pages-limitextraction.llmwiki compile --no-pages-limitremoves the requested 3–8 concepts per source, so a source yields the pages it supports.- Viewer refresh and port.
llmwiki viewrebuilds its snapshot in the background after a compile, stops refreshing cleanly on close, and readsLLMWIKI_VIEW_PORTwhen--portis not given. - Correctness fixes. Wikilink markup in an extracted concept name no longer creates a duplicate page, and GPT-6 Luna models work with the OpenAI provider. MCP wiki tools keep progress output off the stdio protocol stream, and a named pipe in
sources/no longer hangs freshness checks. - Prompt version
v7. Every newly generated page records prompt versionv7, including pages from default runs. Unchanged sources are not recompiled by upgrading alone.
npm install --global llm-wiki-compiler
The supporting packages are pinned to the same version and install with it. The MCP Registry entry io.github.atomicstrata/llmwiki is published for 1.5.0 after the npm release. Existing projects need no migration; read the upgrade notes for the prompt-version change and the opt-in Converse dependencies.
Released 1.4.2: Fixes llmwiki compile with the codex-agent provider and carries the 1.4.0 features: answer review before publishing, batch approval, extraction reuse, page-level answer citations, and the split-package SDK with matching supporting dependencies.
Experimental artifact types and workflow statuses can require source updates; read the SDK upgrade notes before upgrading from 1.3. See the 1.4.0 release summary and changelog for migration details, contributor credits, and limits.
Released 1.3.0:
- Four viewer themes, with Scientific Clay as the new default and Minimal following your system setting. Saved light/dark preferences migrate to Nebula Light/Dark.
- Recursive source discovery, literal path exclusions, and explicit project-specific compile instructions.
- Bounded binary storage for larger embedding indexes, configurable completion budgets, and protected gateway request extensions.
- OrcaRouter embedding routing, retryable rule-language changes, and a public test type-check ratchet.
See the 1.3.0 release summary and changelog for migration details, contributor credits, and limits.
Released 1.2.0:
- Profile-aware viewer with document categories, record relationships, provenance, source previews, verified local artifact access, responsive navigation and fitted graphs.
- Additional providers and OpenAI-compatible reasoning-model controls, independent embedding configuration, and SDK embedding opt-out.
- Optional Sources sections and additive SDK system policies with prompt-change invalidation and per-page provenance.
- Shared-page recovery after source removal, citation line-list fixes, and conservative abbreviated-wikilink repair.
Windows path fixes are included, but native Windows validation and CI remain outstanding. Full filesystem-confinement guarantees on Windows remain unverified. Use a supported Linux environment for untrusted projects; see #217.
Released 1.1.0:
- Template distribution ecosystem: publishers author signed, offline distributions with
template publish init | add | build | rotate | revoke(Ed25519 signing, key rotation, package revocation) and verify them withtemplate publish verify. - Consumers configure explicitly trusted template taps, discover and inspect signed catalogs, and install or update templates with continuity, revocation, and compatibility checks under lock.
llmwiki statuscommand: a readable snapshot of page and source counts, last compile, stale and orphaned pages, pending changes, the review queue, active profile, and state-file health.
Released 1.0.0:
- Configurable Lifecycle Profiles across CLI, SDK, MCP, viewer, context, lint, status, export, and profile-aware OKF exchange.
- Built-in
autosciandnewsroomtemplates, typed workflows and actions, first-class artifacts, typed relations and runtime lifecycle gates, plus a hardened first-party connector substrate with Crossref. - Typed-page semantic search and retrieval controls, batch embeddings, parallel compile, and fail-closed state recovery.
See CHANGELOG.md for release history.
Companion: Atomic Memory
llmwiki and Atomic Memory are complementary open context infrastructure:
- llmwiki compiles source material into durable, inspectable knowledge.
- Atomic Memory gives agents runtime memory that is searchable, scoped, correctable, and inspectable.
Use them independently or together. The @atomicmemory/llmwiki bridge imports llmwiki export --target json --project-id <id> as durable memory records.
Contributing
Contributions are welcome. If llmwiki is missing something you need, open an issue or PR and describe the workflow you are trying to support - need-driven improvements are often the best ones. If you want to contribute more generally, roadmap items are a good place to start. For larger changes to core compile, review, import/export, or retrieval semantics, please start with an issue or design discussion so we can align on the contract first.
Before committing code changes, run:
npx tsc --noEmit
npm run build
npm test
npm run fallow:ci
See CONTRIBUTING.md.
License
MIT
Disclaimer
No LLMs were harmed in the making of this repo.
