openclay-agent 2.0.0

Turn intention into infrastructure. Local-first AI agent bootstrapper.

🏺 OpenClay — turn intention into infrastructure.

You describe what you want. OpenClay reads your machine, builds a local AI stack, and starts working. No config files. No copy-paste. Everything it learns about you stays on your computer, in files you own.

pip install openclay-agent

Or from source:

git clone https://github.com/openclay1/OpenClay.git && cd OpenClay && pip3 install -r requirements.txt && python3 app.py

Your hardware	What OpenClay runs
32 GB+ RAM, Apple Silicon / 6 GB+ VRAM	llama3:8b — full reasoning
16 GB+ RAM, Apple Silicon / CUDA	qwen2.5:7b — fast + capable
16 GB+ RAM, Intel, no discrete GPU	qwen2.5:3b-instruct-q4_K_M — lean workhorse
8 GB+ RAM	qwen2.5:1.5b — lightweight
Under 8 GB	Template-only mode — no model needed
Any machine + API key	Claude, GPT-4, etc. via config

---

What it does today

• Single input, single button. Type an intention, hit Go.
• Hardware detection → model selection → silent install. You don't pick a model. It does.
• LLM Wiki. Karpathy-pattern knowledge base. Ingest files, query your wiki, lint for health. Local markdown, yours forever.
• Tweet drafting and posting. Draft from intention, review, post — or post directly.
• Persistent memory. AGENTS.md tracks what works, what fails, your preferences. The agent reads it before every action.

How it works

intention → hardware scan → model install → execution → wiki memory

Everything runs locally through Ollama. No data leaves your machine.

Wiki

OpenClay maintains a local wiki that compounds over time. Drop a file in raw/, type ingest filename, and the LLM breaks it into concept pages, entity pages, and source summaries. Query it later. Lint it for contradictions.

raw/              ← your files, immutable, LLM never writes here
wiki/
  concepts/       ← idea pages
  entities/       ← people, orgs, tools
  sources/        ← one summary per ingested file
  comparisons/    ← cross-source analysis
  index.md        ← auto-generated catalog
  log.md          ← append-only activity record

The wiki is not for you. It's for the agent — so it can act with consistency.

Configuration

cp .env.example .env

Most things work with zero config. Add Twitter keys only if you want to post.

Project structure

app.py              — entry point
panel.py            — browser UI (Gradio)
agent_backend.py    — switchable LLM backend
wiki_engine.py      — wiki operations: ingest, query, lint
memory.py           — AGENTS.md persistent memory
twitter_post.py     — tweet posting (Tweepy)
introspect.py       — hardware detection
theme.css           — design system
openclay.md         — wiki schema
wiki/               — agent's compounding memory
raw/                — immutable source documents

Every module stays under 300 lines.

Principles

1. Act, don't ask. If the agent can do it, it does it.
2. Local-first. No cloud dependency for core function.
3. Files over apps. Markdown you own, not data locked in a platform.
4. Install what's missing. Silent. No questions.

License

MIT — it's yours.

github.com/openclay1/OpenClay