agent-wiki

A Karpathy-style knowledge base that any LLM agent can read and maintain.

You drop sources, the AI compiles a wiki, and knowledge compounds over time.

Prerequisites

Requires: bash, git. Optional: rg (ripgrep) for faster search.

Quick start

Open in your agent (Claude Code, Codex, Cursor, Hermes) and say:

That's it. The agent reads the wiki, briefs you, and you're working.

How to use

When you have a source to add

The agent saves it to raw/, reads it, asks you socratic questions about what it means, compiles it into the wiki, and cross-links with existing pages.

Weekly health check

Finds contradictions, orphan pages, stale content, dead links, unsourced claims, and suggests new pages to fill gaps.

All commands

CLI commands (the engine behind the skill)

You don't usually call wikictl directly — the skill commands call it for you.

Two modes

Mode 1: In-Wiki

Open your agent in this repo. Work directly in raw/, wiki/, outputs/. The repo is your wiki.

Mode 2: Second-Brain

Open your agent in another repo (client code, project, etc.). Agent-wiki lives alongside as shared memory. Your code stays in the code repo. Durable knowledge goes into agent-wiki.

Both modes use the same 5 commands.

MCP (optional)

If your agent speaks MCP: