Guides
A hosted LLM wiki your agents maintain
An LLM wiki is a folder of Markdown pages that an AI agent writes and keeps current from your raw sources, instead of you. Cozy hosts that wiki for you. Every account gets a private workspace and one remote MCP endpoint, so any agent on any machine can ingest, query and lint the same pages without a local vault, tunnel or sync layer.
What the LLM wiki pattern is
In April 2026 Andrej Karpathy published the LLM Wiki idea file. It describes three layers and three operations:
- Raw sources: the original articles, transcripts, notes and documents, kept unchanged.
- The wiki: Markdown pages the LLM owns and rewrites, with an
index.mdmap and alog.mdof changes. - The schema: a file such as
CLAUDE.mdorAGENTS.mdthat tells the agent how to organize the wiki. - Ingest, query, lint: add a source and update every page it touches; answer from the wiki; periodically find contradictions, stale claims and orphan pages.
His key observation is that people abandon wikis because the maintenance grows faster than the value. An agent can do the bookkeeping.
Why the local-vault version breaks down
Most tutorials build the pattern as an Obsidian vault on one laptop, plus Claude Code. That works for one person, one machine and one agent. It gets hard when:
- You want the same wiki from claude.ai on your phone, ChatGPT, or a second computer. Obsidian has no cloud version, and community MCP servers need the desktop app running or a tunnel such as Tailscale Funnel or Cloudflare Tunnel.
- Two agents write at once. Many community Obsidian MCP servers overwrite files, so the last write wins.
- A claim goes stale because its source changed, and nothing tells you.
- You want an agent to read your research folder but not your personal notes.
How Cozy maps to the pattern
| Karpathy's layer or operation | In Cozy |
|---|---|
| Raw sources | sources/ folder, filled by Connectors (Granola, Gmail, Google Drive, Notion, Slack, Linear and more), uploads, or the ingest API, with stable IDs, URLs and timestamps |
| The wiki | wiki/ folder of plain Markdown with an index.md map, wikilinks, backlinks, aliases and typed relations |
| The schema | Editable SKILL.md files in skills/, served to agents as MCP prompts (organize, recall, maintain, dream) |
| Ingest | cozy_ingest saves evidence; the agent then updates the pages it touches |
| Query | cozy_context returns a page, its typed neighbors with excerpts, its evidence and warnings in one call |
| Lint | cozy_health: broken links, duplicates, stale pages, conflicts, orphans, unsourced pages, a score and prioritized fixes |
| log.md | Revision history (last 50 revisions per file, restorable) and cozy_changes since any timestamp |
What a hosted version adds
- One endpoint for every agent. Claude, Claude Code, Codex, ChatGPT and other MCP clients connect to
https://cozy.cloud/mcpand sign in with OAuth. One URL reaches all your workspaces. Scripts and headless agents can use a per-workspace URL with a bearer credential instead. - Safe concurrent writes. Every edit carries the revision it was based on. A conflicting edit is rejected rather than silently overwriting another agent's work.
- Provenance and staleness. Pages cite files in
sources/. When a cited source changes after the page was last edited, Health flags the page as stale. Disagreements are recorded asCONFLICT:lines. - Scoped access. An agent never gets more access than you have in a workspace, capped by the scopes you approved. Connected agents can be disconnected under Connectors → Agents. Workspace credentials for scripts carry read, wiki-write or source-write scopes and optional folder prefixes, expire within 90 days and can be revoked.
- Dreaming. A consolidation pass, inspired by Anthropic's agent-memory research, that your own agent runs between tasks: check claims against sources, merge duplicates, write typed relations and keep the index short.
Cozy runs no model itself. You bring the agent, and the wiki stays plain files you can read and take with you.
Set it up
- Sign in at cozy.cloud. Your account gets My wiki automatically.
- Add
https://cozy.cloud/mcpto Claude, Claude Code, Codex, ChatGPT or Cursor as a remote MCP server, then approve access on Cozy's consent page. See the Cozy MCP server guide. - Ask the agent to call
cozy_workspacesif you use more than one workspace. Tools use My wiki unless the agent passes aworkspace. - Connect a source or push some notes, then ask your agent to run the organize prompt and build
wiki/index.md.
Frequently asked questions
What is an LLM wiki?
An LLM wiki is a collection of Markdown pages that an AI agent writes and maintains from your raw sources. You add sources and ask questions; the agent updates pages, links them and periodically lints the wiki for contradictions and stale claims. Andrej Karpathy described the pattern in April 2026.
Do I need Obsidian to run an LLM wiki?
No. Obsidian is a popular viewer for a local vault, but the pattern only needs Markdown files and an agent. Cozy hosts the files and exposes them over MCP, so you can use the wiki from any MCP client without a local vault.
How is an LLM wiki different from RAG?
RAG retrieves raw chunks at question time. An LLM wiki compiles sources into maintained pages ahead of time, so answers come from curated, linked, cited pages. Cozy uses lexical search over those pages and needs no embeddings.
Can several agents share one LLM wiki?
Yes. In Cozy every write includes the revision it was based on, so concurrent edits conflict instead of overwriting each other. Each agent signs in separately, with its own scopes, and can be disconnected on its own.
Give your agents a memory worth keeping
Cozy gives every account a private wiki that Claude, Codex, ChatGPT and other MCP clients can read and maintain.
Start your wiki