Texo

Your thoughts are a folder. Your own AI weaves them into a wiki.

Texo turns passing thoughts into an interlinked personal wiki — plain markdown, compiled by your own AI agent, owned entirely by you. Built on Andrej Karpathy's LLM Wiki pattern.

It's two pieces — no account, no server, nothing to sign up for:

• texo-web/ — a zero-dependency local web client. Capture thoughts in your browser; browse and search your wiki. Writes plain markdown to a folder you own.
• texo-skill/ — the compiler. Your own command-line agent (Claude Code, Codex, …) reads the raw fragments and weaves them into concept / opinion / decision / entity pages, with aliases and backlinks.

The point: turn scattered notes into a personal wiki that's actually useful to read back.

How it works

 capture (web)              your folder (markdown)          compile (your agent)
  type in browser  ──▶  raw/   ── immutable originals  ──▶   reads raw/
                        wiki/   ── woven by your agent  ◀──   writes wiki/
                        schema.md ── the rules both obey

Three layers, per the LLM Wiki pattern. Write ownership is disjoint — the client only writes raw/, the compiler only writes wiki/ — so they never conflict.

Quickstart

1. Capture — run the web client (needs Node 18+):

git clone https://github.com/partyfly/texo && cd texo
node texo-web/server.mjs          # → http://127.0.0.1:4321

First run scaffolds your library at ~/Texo (override with TEXO_DIR). Open the page and jot a few thoughts.

2. Compile — let your own agent install itself. There is no installer script: INSTALL.md is a spec your agent executes, adapting to your OS, your skill directory, and whatever engine you use. Paste into Claude Code (or any CLI agent):

Install Texo by following INSTALL.md in https://github.com/partyfly/texo.

It installs the skills, copies the runtime pieces verbatim, asks how you want compilation triggered, verifies everything against a checklist, and records what it did in ~/.texo/install-manifest.md so uninstall/upgrade always has a paper trail.

Or inside Claude Code: /plugin marketplace add partyfly/texo → /plugin install texo-skill@texo.

Three ways to trigger compilation

Same compile core, bring your own engine (claude / codex / any CLI agent), pick one — the texo skill's watch operation walks you through it:

• Background loop (recommended) — an OS timer (launchd / cron / systemd) wakes your agent only when there are new entries (a zero-token shell precheck gates it). Lowest running cost, fully hands-off, runs even with nothing open.
• On demand — click Compile in the web client, or tell your agent "compile my Texo".
• Agent built-in scheduled task — zero install; your agent's own scheduler (Claude Code scheduled-tasks, WorkBuddy automation tasks, QwenWork desktop 定时任务, Trae SOLO tasks, …) runs it periodically.

Recall — your wiki, read back into your work

Compilation is only half the loop. The other half is recall: when one of your agents is doing real work — writing, deciding, building — it should reach into what you've already concluded instead of starting from blank.

• texo skill's recall op (pull) — mid-task, an agent looks up your recorded opinions, decisions, and articles on the topic and folds them into its work, read-only. It works like tool search: a small matcher (recall.py) ranks pages by title + aliases (each page's one-line description is the relevance handle), and the agent expands only the top matches — so document recall reuses the same instinct the model already has for tool recall, instead of reading the whole library.
• SessionStart hook (push) — ~/.texo/recall-hint.sh injects a tiny skeleton of what you've recorded (categories + page titles) at the start of every session, so the agent knows there's something worth pulling. It reads only the index, never page bodies, and stays silent when there's no library.

The hook is registered in ~/.claude/settings.json, so it surfaces across your projects. It's plain, fast, and reversible — to turn it off, drop the SessionStart entry pointing to recall-hint.sh.

File over app

• Your data is a folder of plain markdown. Opens in Obsidian. Outlives this tool.
• The AI is yours — compilation runs locally, with your agent and your account. There is no Texo server.
• Capture, browse, and search never need a network.

What's here

texo-web/      local web capture client (zero deps; capture · browse · search)
texo-skill/    one `texo` skill (SKILL.md router + references/ per operation) + schema templates
               — runtime scripts ship inside the skills that own them
               (watch/runtime/*, recall/scripts/recall-hint.sh), copied verbatim at install
INSTALL.md     the install spec your agent executes (no installer script)

Together they're the whole loop — capture in the browser, compile with your agent, read it back — running entirely on your machine.

License

MIT. The schema is append-only by design — anything that reads or writes a Texo folder today should still work years from now.