Molly OS Documentation

Introduction

Molly is a sovereign multi-agent orchestrator. Rather than sending every request to a single external model, Molly manages a fleet of task-specialized adapters over one quantized base model, and routes each request to the cheapest, most private, most capable place it can run.

The guiding principle is your intelligence, your hardware, your rules. Inference runs on-device by default and only escalates outward when your policy explicitly allows it. Nothing leaves your control unless you say so.

Who Molly is for

• Teams that need data residency and cannot ship sensitive data to third-party clouds.
• Builders who want a private AI specialist for their own domain without hiring ML engineers.
• Developers who want an OpenAI-compatible API they can point at existing code.

Quickstart

Molly exposes an OpenAI-compatible API, so if you already use the OpenAI SDK you can switch by changing the base URL and key.

First call

# Python — OpenAI-compatible client
from openai import OpenAI

client = OpenAI(
    base_url="https://api.iamolly.ai/v1",
    api_key="YOUR_MOLLY_API_KEY",
)

resp = client.chat.completions.create(
    model="molly-orchestrator",
    messages=[{"role": "user",
               "content": "Summarize this contract clause."}],
)
print(resp.choices[0].message.content)

The CEO Multi-Agent Orchestrator

At the center of Molly is the CEO orchestrator — a coordinating agent that interprets each request, decomposes it into subtasks, and delegates to the right specialists. It behaves like a chief executive assigning work to a team of experts.

What the orchestrator does

• Understands intent and breaks complex tasks into steps.
• Selects specialists — chooses the LoRA adapter best suited to each subtask.
• Chooses the tier — decides where execution happens under your routing policy.
• Composes results from multiple specialists into a single coherent answer.

Specialists (LoRA Adapters)

Instead of running many large independent models, Molly loads one quantized base model and swaps in lightweight LoRA adapters — each a compact specialist trained for a specific domain or task.

4-Tier Routing (T0–T3)

Every request is routed through a four-tier privacy hierarchy. Local execution is the default, and escalation to higher tiers is policy-gated — Molly will not reach for an external API unless your policy permits it.

Sovereign & Federated Execution

Molly is sovereign: intelligence lives on your hardware and operates under your rules. When multiple nodes collaborate, Molly uses federated learning — nodes share adapter deltas, never raw data.

How federation works

• Each node trains and improves specialists on its own local data.
• Only adapter deltas (the learned weights) are exchanged — raw data stays put.
• Improvements aggregate across the fleet without centralizing sensitive information.

Continuous Self-Improvement Loop

Molly improves over time through a closed loop: traces → eval → distilled adapters.

Multimodal

Molly's orchestration model is multimodal — specialists can be composed across different input and output modalities, coordinated by the same CEO orchestrator and governed by the same routing policy.

Plug-and-Play Training

Build your own personal AI without being an engineer. Point Molly at your data source and it will auto-train, evaluate, and deploy a specialist for you — no ML pipeline to build, no infrastructure to wire up.

The workflow

What to expect

• No ML engineering required — the train → eval → deploy loop is automated.
• Evaluation is handled by a neutral judge, so specialists are gated on measured quality.
• New specialists join the fleet and improve over time via the self-improvement loop.

Privacy & Security

On-device by default

Inference defaults to T0 on-device. External tiers are only reached when your policy explicitly allows it, so sensitive workloads stay local unless you choose otherwise.

Data residency

You control where execution and training happen across the T0–T3 hierarchy. In federated setups, nodes exchange adapter deltas, never raw data, keeping information within its home boundary.

Merkle-tree provenance

Molly maintains Merkle-tree provenance, giving a verifiable, tamper-evident record of how models and adapters were produced and changed.

Pricing Summary

Because Molly runs local-first, much of your inference can happen on your own hardware, with external tiers used only when your policy allows. For current plans, token packs, and tier details, see the pricing page.