# LLM Wiki Compiler

Compiles raw sources into a cited, interlinked markdown wiki; CLI, SDK, and MCP server.

llmwiki turns raw files and URLs into typed, citation-linked markdown pages that agents and people can query, lint and export. The same package ships a CLI, a TypeScript SDK and a stdio MCP server (llmwiki serve). A JSON profile and an instructions file change what the wiki records and how it is written.

- Capabilities: adaptable, operable
- Vertical: notes-knowledge
- Shape: product
- Form factor: cli
- Execution model: hybrid
- Provenance: unknown (recorded, never ranked)
- Repo: https://github.com/atomicstrata/llm-wiki-compiler (MIT), 2242 stars, last commit 2026-10-30
- Last activity: 2026-10-30
- Page: apps/atomicstrata--llm-wiki-compiler.html
- JSON: apps/atomicstrata--llm-wiki-compiler.json

## Use it when

- you want a durable, cited knowledge base your agent can query instead of re-reading raw files
- you run a research or editorial project and want typed records with review gates

## Gotchas

- Needs Node.js 24 or newer
- compile and query call a paid model unless you point it at a local one
- The README calls it early software
- Repo moved from atomicmemory to atomicstrata
- Each provider has its own key; Ollama, the local Claude Code login and Codex need none

## Editions

- Free: Free, open source
  - source: https://github.com/atomicstrata/llm-wiki-compiler
  - installable: npm install --global llm-wiki-compiler

## Adaptable: make it yours

Effort: config-only (8 things to edit)

A validated .llmwiki/profile.json declares entity types, fields, relations, lifecycles, artifacts, workflows and gates; llmwiki template init autosci or newsroom scaffolds one. llmwiki compile --instructions ./SOUL.md adds writing guidance and .llmwiki/config.json sets review policy. LLMWIKI_PROVIDER, LLMWIKI_MODEL and LLMWIKI_OUTPUT_LANG choose provider, model and output language.

- Env vars: LLMWIKI_PROVIDER, LLMWIKI_MODEL, LLMWIKI_OUTPUT_LANG, ANTHROPIC_API_KEY, OPENAI_API_KEY
- Config files: .llmwiki/profile.json, .llmwiki/config.json
- Prompt files: SOUL.md (any file passed to --instructions)
- Accounts you'll need: a model provider account or a local Claude Code, Codex or Ollama setup

## Operable: what an agent can drive

Operability: official. Interface: mcp at https://github.com/atomicstrata/llm-wiki-compiler/blob/main/docs/guides/mcp-agent-integration.mdx (published by the builder, local).

### Reach (partial coverage)

- Ingest URLs and files as sources: reachable. MCP tool ingest_source; CLI llmwiki ingest
- Compile sources into wiki pages: reachable. MCP tool compile_wiki; needs a model provider key
- Ask grounded questions and save answers: reachable. MCP tool query_wiki with optional save
- Find and read pages: reachable. MCP tools search_pages, read_page
- Build a citation-aware evidence pack: reachable. MCP tool get_context_pack; works without a provider key
- Lint structure, citations and freshness: reachable. MCP tools lint_wiki, wiki_status
- Score wiki quality: reachable. MCP tool run_eval (fast suite needs no key)
- Exchange knowledge as Open Knowledge Format: reachable. MCP tools export_okf, import_okf; import is staged for review
- Approve or reject held review candidates: not reachable. llmwiki review approve is CLI only; MCP tools list has no review tool and run_workflow_action cannot approve human gates
- Remove a source, repair stale pages, install profile templates: not reachable. llmwiki rm, refresh --stale and template init are CLI commands; none appear in the MCP tool table

## Code stack

- Languages: TypeScript
- Frameworks: none
- Package manager: npm
- Database: none
- Hosting target: local

## Model stack

- Providers: Anthropic, OpenAI-compatible, Ollama, GitHub Copilot, Claude Agent SDK, OpenAI Codex CLI, MiniMax, Converse gateways, OrcaRouter, Atlas Cloud
- Models: none
- Keys you need (names only): none
- Prompts live in: any file passed with llmwiki compile --instructions
- Tools: ingest_source, compile_wiki, query_wiki, search_pages, read_page, lint_wiki, wiki_status, get_context_pack, run_eval, export_okf, import_okf, verify_artifact, list_workflow_actions, describe_workflow_action, run_workflow_action, workflow_run_status

## Skills

bundled: A SKILL.md shipped in the npm package tells an agent to call wiki_status, then get_context_pack, before generating.
- Paths: skills/llmwiki/SKILL.md
- Install: mkdir -p ~/.claude/skills && ln -s "$(npm root -g)/llm-wiki-compiler/skills/llmwiki" ~/.claude/skills/llmwiki

## Evidence and confidence

- formFactor: README: npm global CLI llmwiki, SDK createWiki, local viewer (confidence 0.9)
- executionModel: README: two-phase LLM pipeline; lint, status and fast eval need no provider (confidence 0.85)
- operability: src/mcp registers 16 tools, matching docs/guides/mcp-agent-integration.mdx; the npm package's repository field points to this repo (confidence 0.9)
- effort: docs/configuration/profiles.mdx gives the profile shape (schemaVersion, entities, relations, lifecycles, workflows); src/profile/load.ts fails closed on a broken file; docs/configuration/review-policy.mdx covers config.json; docs/cli/compile.mdx covers --instructions; docs/configuration/environment-variables.mdx lists LLMWIKI_* vars (confidence 0.9)
- editions: MIT repo and npm package; no pricing page on atomicstrata.ai or the docs site (confidence 0.8)
- active: latest commit 2026-10-30, release v1.5.0 on 2026-10-19
- skills: docs/guides/mcp-agent-integration.mdx, Agent skill section, gives the symlink command; package.json files includes skills/llmwiki/
- reach:  (confidence 0.9)

## Sources

- https://llmwiki.atomicstrata.ai (sourcing:notes-knowledge:community)
- https://github.com/atomicstrata/llm-wiki-compiler (analysis:notes-knowledge)