agent-wiki-kit

One engine for every LLM wiki you'll ever run. A wiki is a folder of Markdown files with frontmatter plus a wiki.yaml registry — wikikit makes it lintable (contracts, links, freshness), publishable (llms.txt, llms-full.txt, index.json, plain no-JS HTML), and consumable by any agent over MCP. Zero dependencies: Python 3 stdlib only.

The pattern follows the Karpathy LLM-wiki idea and the agentwikis.com serving model: knowledge is compiled once and kept current, not re-derived per query. Compilation is agent work (the wiki-ingest skill / wikikit ingest); the engine itself is deterministic.

See it live: examples/corporate-brain-patterns is a real wiki built with wikikit — read INDEX.md for the source pages, or the generated _site/ for what wikikit build emits (llms.txt, llms-full.txt, index.json, HTML).

Why not just a vector DB / RAG over raw docs?

Raw-text retrieval re-derives an answer from whatever chunks happen to match a query, every single time — no page ever gets better, corrections don't stick, and nothing tells you a source went stale. wikikit's bet: a small, governed, continuously-curated wiki that an agent reads directly beats re-deriving from an ever-growing pile of raw source documents, both on accuracy and on maintenance cost. See reference/curated-vs-raw-retrieval for the evidence. It isn't a replacement for a vector DB on unstructured search — it's for the slice of knowledge worth compiling once and keeping current, the same slice a human onboarding doc or runbook already covers.

Quick start

pip install -e .            # or: pip install wikikit, once published

# scaffold a wiki (constitution + registry + meta)
wikikit init ~/wikis/my-domain --name my-domain

# compile a source into it (uses the claude CLI; or use skills/wiki-ingest interactively)
wikikit ingest ~/wikis/my-domain ~/Downloads/some-research.md

# health
wikikit lint ~/wikis/my-domain      # errors exit 1 — CI-able
wikikit status ~/wikis/my-domain

# publish for agents + humans
wikikit build ~/wikis/my-domain     # -> _site/llms.txt, index.json, HTML

# serve to any MCP client
wikikit serve ~/wikis/my-domain

No install? Every command also runs as python3 wikikit_cli.py <cmd> ... from a clone of this repo.

Register with Claude Code / Claude Desktop:

{"mcpServers": {"my-domain-wiki": {
  "command": "wikikit",
  "args": ["serve", "/path/to/wikis/my-domain"]}}}

MCP tools exposed: wiki_list_pages, wiki_read_page, wiki_read_section (cheap targeted reads), wiki_search, wiki_recent. Pages are re-read on each call, so edits are live without restarting.

Page contract

---
title: GitLab's Public Handbook
type: entity          # from wiki.yaml page_types
status: live          # live | draft | deprecated | planned
last_verified: 2026-07-07   # a claim: checked against reality on this date
summary: One line for llms.txt listings and search results.
sources: [where this came from]
tags: [a, b]
---

[[wikilinks]] are first-class (lint checks them both in [[slug]] and relative .md link form). Dirs starting with _ (_raw, _meta, _site) are operational, never served.

Lint rules

MALFORMED, MISSING_FIELD, BAD_STATUS, BAD_DATE, UNKNOWN_TYPE, BROKEN_LINK (errors — exit 1) · STALE, ORPHAN, NO_SUMMARY (warnings).

Layout

wikikit/            engine (frontmatter, wiki model, lint, search, build, mcp_server, ingest, cli)
templates/          scaffolding used by `init` (CONSTITUTION.md, wiki.yaml)
skills/wiki-ingest/ Claude Code skill: how to compile sources into a wiki
examples/corporate-brain-patterns/  working demo — public case studies in compiled organizational knowledge (lints clean, serves over MCP)

Instance profiles (roadmap)

The same engine backs different products by choosing page types + ingest sources:

• Market/competitor wiki — entities + synthesis + weekly signals, for tracking a competitive landscape.
• Key-account wiki — snapshot, buying map, tech stack, angle, and signals per account.
• Data-asset wiki — tables, semantic models, pipelines, and agents, feeding downstream analytics agents.
• Ops wiki — runbooks + known-error DB.