milna-os 1.0.0

BYOK multi-provider AI runtime. Your keys, your compute, your output.

Milna OS

Powered by Kilonova

A multi-model AI terminal assistant that runs your keys, your compute, your output — no subscriptions, no data leaving your machine without your knowledge.

Milna fires your prompt across all active AI legs in parallel, reconciles their outputs into a single synthesized response, and lets you build persistent knowledge bases (DOTs) from your own documents.

---

What it does

• Multi-leg parallel inference — Anthropic, OpenAI, Gemini, and local Ollama run simultaneously. Responses are reconciled, not averaged.
• DOT (Document of Truth) — Ingest a folder of PDFs into a persistent, structured knowledge base. Load it as a context anchor for any conversation.
• BYOK — Bring your own API keys. They live in your OS credential store, never on disk in plaintext.
• Free local fallback — Ollama on localhost runs for free when cloud legs are unavailable or too expensive.
• Session logging — Every turn is written to disk as markdown. Full audit trail.
• No subscriptions. No compute surcharge. Remember the days when you bought something and you just... owned it? This is a tool. I have a hammer and I don't pay by the nail. Your output is yours — never gated, never watermarked.

---

Requirements

• Python 3.11+
• Ollama (optional but recommended — free local inference)
• At least one API key: Anthropic, OpenAI, or Gemini

---

Install

pip install milna-os

Python 3.11+ required. On Windows, keys are stored in Windows Credential Manager. On Mac/Linux, the OS keychain is used — never on disk in plaintext.

---

Run

milna

First run: Milna will prompt you to add an API key.

To add or update a key at any time:

milna --add-key

---

Source

git clone https://github.com/MilnaOS/milna-os.git
cd milna-os
pip install -e .

---

Commands

Command	Description
/load <path>	Load a PDF, .md, or .txt file into conversation context
/load <folder>	Pick from files in a folder
/fetch <url>	Fetch a web page into conversation context
/webseed <topic> <url>	Extract entities from a web page into a DOT
/webseed <topic> <url> --follow	Crawl linked pages (depth 1, up to 20 by default)
/webseed <topic> <url> --depth <n>	BFS crawl N levels deep
/webseed <topic> <url> --force	Re-fetch all URLs, ignoring saved crawl state
/ingest <folder>	Build a persistent knowledge base (DOT) from PDFs
/ingest <folder> --deep	Deep distillation — richer extraction, higher cost
/ingest <folder> --topic <name>	Custom topic name for the DOT
/ingest --rebuild <topic>	Re-extract entity records from existing distilled docs
/ingest --rebuild <topic> --force	Wipe and re-extract all records
/ingest --rebuild <topic> --force --deep	Wipe, re-extract with deep distillation
/kb	List available knowledge bases
/kb <topic>	Load a DOT as conversation context anchor
/deep	Enable deep distillation for the next message
/mode	Show current processing mode
/mode <name>	Switch mode: potato / economy / standard / blade
/legs	Show active legs and their status
/keys	Add or update API keys
/ollama	List available Ollama models
/ollama <model>	Switch active Ollama model
/reset	Clear conversation history
/help	Show command list
exit / quit	Shutdown

---

Modes

Mode	Isolates per leg	Use when
potato	1	Free local-only, Ollama routes everything
economy	1	Light cloud usage, cost-conscious
standard	2	Default — balanced quality and cost
blade	3	Maximum quality, highest cost

---

Building a DOT

A DOT (Document of Truth) is a structured knowledge base built from source documents. Once built, it loads instantly as a permanent context anchor — models treat it as verified reference material, not something to hallucinate around.

# Ingest a folder of PDFs
/ingest D:/my-documents/research --topic my_research

# Load the DOT into conversation
/kb my_research

# Ask questions grounded in your documents
> summarize the key findings from the research

DOTs are stored in kb/<topic>/ and persist between sessions. Incremental ingest adds new documents without re-processing existing ones.

---

Project structure

milna.py                    — CLI entry point
kilonova/
  core/
    engine.py               — ConversationEngine: parallel firing + reconciler
    legs.py                 — Leg registry, key inventory, provider wiring
    reconciler.py           — Multi-role synthesis
    mode.py                 — Mode controller (potato/economy/standard/blade)
    session_log.py          — Turn logging to disk
    providers/
      anthropic_provider.py
      openai_provider.py
      gemini_provider.py
      ollama_provider.py
  tools/
    loader.py               — PDF / .md / .txt extraction
    ingest.py               — DOT build pipeline
kb/                         — Built knowledge bases (gitignored)
session_logs/               — Turn transcripts (gitignored)
prompts/                    — Loadable system prompt files

---

Local-only operation

Run fully offline with Ollama:

ollama serve
ollama pull qwen2.5:3b
python milna.py --mode potato

No API keys required. No data leaves your machine.

---

License

Proprietary — source available for inspection. See LICENSE file for full terms. Licensing inquiries: [e-mail]

---

BYOK — your keys, your compute, your output.