Metadata-Version: 2.4
Name: holarch
Version: 0.2.0
Summary: One console for your AI work: cheapest path first, your notes before any model, works offline. CLI + MCP.
License: MIT
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Provides-Extra: intent
Requires-Dist: scikit-learn>=1.3; extra == "intent"
Requires-Dist: numpy; extra == "intent"
Dynamic: license-file

# holarch

**One console for your AI work. It uses the cheapest path that can do the job, and it keeps working when the internet or your cloud credits don't.**

## Install
```bash
pipx install holarch            # or: pip install holarch ; add [intent] for the free intent router
holarch doctor                  # what is set up
holarch index ~/notes           # your notes become memory (public/ personal/ private/ folders = rings)
holarch "what did I decide about the launch?"
claude mcp add holarch -- holarch mcp      # use it from Claude Code (holarch ide prints Cursor / VS Code config)
```
Local model: install [Ollama](https://ollama.com) and `ollama pull qwen3:1.7b` (and `nomic-embed-text` for better note search).
Cloud: set any of `DEEPSEEK_API_KEY`, `OPENAI_API_KEY`, `OPENROUTER_API_KEY`, `GEMINI_API_KEY`, or a holarch AI Credits key.
Keys stay in your environment; holarch never stores or prints them.

## How it answers
1. **Your tools first.** Ask for something by name and holarch runs it directly. No model call.
2. **Your notes next.** It searches your own files and notes before it asks any model.
3. **A small model on your machine.** It writes the answer from what it found. No internet needed.
4. **A long-context brain for planning.** Big "what's the plan / summarize everything" questions go to your NotebookLM notebooks. The answer comes back with citations and is saved into your notes, so next time it's local.
5. **Cloud models last.** Only when they're up and worth it. holarch notices a dead or out-of-credit provider and stops calling it.

## Use it from anywhere
- **Terminal:** `holarch`, plain language or commands.
- **MCP server:** for Claude Code, Cursor, VS Code and ChatGPT. Ships with notes, receipts and health tools; add more as plugins (the `holarch.tools` entry point).
- **IDE:** through the MCP config, plus the terminal.

## Measured (not estimated)
| | |
|---|---|
| Offline, it picked the right tool | 4 of 4, and matched cloud answer quality (0.875 vs 0.875) |
| A dead provider costs | under 1 ms instead of 0.7–2.7 s per wasted call |
| Free intent routing | 81% accurate at $0, ~1 ms (a paid model: 88%) |
| Planning answers from your notebooks | median 40 s, cited, 20 of 20 answered |

## Plans
- **holarch Core:** free. Your keys, your machine, the offline stack, the MCP server.
- **holarch Pro:** the notebook brain set up for you, sync, priority support.
- **holarch AI Credits:** prepaid packs to use cloud models through holarch.
- **Holon setup:** we set holarch up for you or your team on your own notes and workflows. Starts with an onboarding call.

Every plan starts the same way: pick it, pay (or not, for Core), and book your onboarding.
