Questions, answered

No. mpl is a toolchain for running and building open models on your own hardware — part of a larger platform we're building — and we shipped it with the curriculum baked in. You learn by running the real thing, not by watching lessons. It's a tool first; the teaching rides along.

mpl (MindPooL) is a proprietary CLI; the models it runs are open weights (Gemma 4, Qwen3, OLMo 2, Llama) — open models, proprietary toolchain. We've shipped far more than a learner's CLI: mpl is a complete toolchain for sovereign local AI. Beyond run and serve, it covers the full lifecycle (RAG, fine-tuning, quantization, eval), bursts heavy jobs to a GPU you own or rent, stands up a persistent cloud runtime, and clusters many boxes under one operator console — fleet and multi-tenant management included. It runs on certified, silicon-neutral hardware through our partners — AMD-first, Apple Silicon certified, NVIDIA in progress. Installing mpl is free; gated verbs require enrollment. mpl is built for individuals, small teams, and education — the Enterprise runs on its own product line.

No. mpl serves individuals, small teams, and education. Mindpool's enterprise-owned AI infrastructure is a separate product family built from open models, open-source components, and infrastructure in your environment. It starts with Bowline: the open-source intelligence layer for task distribution. Bowline watches model traffic in shadow mode, records every decision it would have made as verifiable evidence, and models the economics of moving each class of work to supply you own or rent — open weights in your VPC included.

mpl runs open models on a pinned, sha256-verified llama.cpp build you own end-to-end — no third-party daemon between you and your weights — and adds the whole lifecycle: RAG, fine-tuning (LoRA/QLoRA), evaluation, quantization, benchmarking, a local coach agent, a live operational dashboard, and lab scaffolds, each tied to a module you run yourself. MLX and vLLM are available as optional accelerator-native engines.

It's a ReAct-loop agent that answers with cited passages from the real curriculum instead of guessing, and can run mpl verbs on your behalf with a confirmation prompt. It remembers durable facts across sessions in a confirm-gated local store (every save needs your y; /memory lists it, /memory clear wipes it). Mid-conversation you can swap models (/model <tag>) or move the session to a rented GPU and back (/cloud <name> / /local) without losing the thread, and context is budgeted automatically so long chats trim the oldest turns instead of erroring out. Memory, tools, and retrieval all stay client-side — even when the model itself is running on a cloud node.

Connecting a wallet isn't a crypto gimmick — it's how Mindpool Labs is built. Mindpool Labs (mindpoollabs.com) has built in Web3 for years, including DefiMentor, our initiative for onboarding people into blockchains. We're bringing that decentralized pedigree to generative and agentic AI. The practical payoff: your wallet proves enrollment without accounts or passwords — mpl signs a challenge locally and only reads an on-chain flag. Enrollment settles in a stablecoin on a public blockchain, which means anyone, anywhere can pay and enroll without a bank account, a card, or a regional payment processor — stable value plus blockchain rails give us global reach the usual sign-up flow can't. It never holds your keys and never writes on-chain. Activation is tied to the machine you log in on, so a shared or copied install won't carry your enrollment to another device — log in once per machine.

Enrollment is mentor-led through DefiMentor. Connect your wallet; if it isn't enrolled yet, talk to your Mentor — they walk you through it. mindpool only reads your status on-chain (Polygon); it never takes custody of funds or keys.

Trust Wallet on Polygon. Use the browser extension or Trust Wallet mobile through WalletConnect. The mpl CLI opens the same Trust Wallet chooser for its one-time signature; --manual remains wallet-agnostic for hardware wallets and command-line signers.

You need GPU compute — but not necessarily a discrete card. What matters is the hardware architecture. Apple Silicon and AMD Strix Halo use a unified-memory design where an integrated GPU shares fast system memory, and both clear the enrollment floor (tier T2) that runs the full curriculum. A separate NVIDIA/CUDA card is one path, not a requirement — it mainly speeds up fine-tuning. Below the floor, mpl doctor --profile tells you before you pay, and local execution of the curriculum is blocked — the supported path is the cloud: mpl setup installs only the cockpit toolkit, and mpl cloud launch runs the curriculum on a box you rent.

No subscription, and no per-token markup. Local hardware is a one-time spend — roughly $1,000–$5,000 — and the weights are free, so once you own the box the marginal cost of running a model is your electricity. Renting instead? You still pay no subscription and no markup to us: mpl cloud reads a live price matrix across AWS, GCP, Azure and OCI, picks the cheapest provider you're connected to (on-demand or spot), bills by the second straight from that provider, and tears the box down on command — so you pay for compute at cost, only while it's running. Installing mpl is free; gated verbs need enrollment.

Yes — and mpl does the renting for you. mpl cloud (beta) launches a GPU box you control, billed by the second: it shows a live price matrix across AWS, GCP, Azure and OCI, picks the cheapest provider you're connected to (on-demand or spot), and tears it down on mpl cloud down. Already have an SSH-able GPU? Connect it with mpl fleet add instead. Cloud is cheaper for light or occasional use and the relief valve for the rare large fine-tune; own the metal once it's your daily driver. A box you launch with mpl cloud launch inherits your enrollment and machine identity at launch, so it's ready to use without a separate login.

No. Inference, RAG, and fine-tuning run entirely on your machine — no frontier API, nothing leaves your machine. The only network call is a periodic on-chain read to verify enrollment (your wallet address, never your data).

macOS (Apple Silicon) and Linux natively; Windows via WSL2. A one-line installer drops a single signed binary on your PATH.

Mostly software engineers. There are three levels: Use (no coding), Automate (any engineer — the agent-harness track, no deep ML math), and Build (the deep end — inference internals and training models from scratch, where real ML background helps). Most people stay in Use and Automate; Build is optional. Lesson text is still being filled in; the lab scaffolds you run are live.