Josue Arcos

Multi Auto Setup (MAS) is an open-source control layer (Apache-2.0) that turns the computers you already own into one coordinated AI system on your own network: it discovers machines, measures what each can run, prepares the right AI environment with your approval, and routes work by explicit rules. Confidential data never leaves your network. WHY: Office PCs can be on up to 77% of the day and idle 61% of that time (NRDC/UCI). A person or small office has a Mac, a Windows laptop, a Linux box, local models, cloud APIs and several AI tools, and today a human is the glue: setting up each machine by hand and sending sensitive data to the cloud to use AI. ARCHITECTURE 1. Discover: one coordinator, nodes join only with owner consent. Printer-style LAN discovery, six-digit one-time PAKE/SPAKE2 code, pinned Ed25519 identities, mutual TLS 1.3, signed messages. Nodes dial out and open no listening ports. REAL test: Mac coordinator + Windows 11 Lenovo; also tested on Linux. 2. Understand: read-only inspection (OS, CPU, RAM, GPU, storage, software, runtimes, models). Each finding is a typed fact with evidence and source; unknowns are recorded, not guessed. 3. Prepare: Setup Builder fits an environment to measured resources and builds a plan from hash-pinned recipes (e.g. a local model via llama.cpp/Ollama, or a harness downloaded from GitHub with SHA-256 verification). One owner approval tied to the plan hash; every operation has an undo. 4. Equip: for any task (coding, setup, CSV analysis, finance, legal...), the Recommendation Engine proposes environment x model x access x machine from verifiable sources (owner subscriptions, GitHub/PyPI/npm packages filtered by license and activity, public benchmarks) and lists what is missing. 5. Coordinate: a deterministic Router picks what runs now by written rules: sensitivity, admission (code-graded exam, 8/10, proof bound to model+environment+suite), cost (local before API before subscription), measured time, load. Confidential work only on owned local machines. Every run records which machine and model answered. Skills, Protocols and Guards (rules in code) keep the model from deciding what it must not. Clients: Claude Desktop, Muse Code, Claude Code (MCP) and Open WebUI (Native Pipe). IMPACT: small and big offices around the world can use AI on client and company data without sending it out (GDPR and similar laws). Small offices without IT (accountants, law firms, clinics) get machines set up without an integrator. Big company offices can run large local models on their own hardware, send heavy jobs to the machines with the most memory or GPU, and use the compute already spread across their macOS, Windows and Linux fleet. Also: rebuild a new PC from another's setup, prepare a whole classroom or lab from one machine, night-shift jobs on idle PCs, lower API bills. Today: secure pairing, inspection, approved setup with undo and harness install work; the multi-machine task board and full routing are early.