Accounts & $12/mo Stripe
Custom auth (bcrypt + JWT + email verification) running on LVLUP’s own backend, backed by DigitalOcean Managed PostgreSQL. Entitlement is a signed JWT cached on-device so offline use keeps working.
AI.LVLUP.NETWORK
LVLUP Agents AI is a local-first, offline-by-default multi-agent platform for macOS. ORION takes a plain-language objective, breaks it into tasks, delegates to worker agents, and finishes the work — with every model, file, and memory running on your own machine. Online providers are opt-in and clearly labeled.
This page describes v1.3.0, the paid public release: accounts, Stripe billing, Online Models, Orion’s Brain, and the rest of the feature set below. See Roadmap.
What’s in the app
Your command center, not just another chat box.
A single macOS app that plans, delegates, executes, and verifies real work — powered by local large-language models running on your own hardware. Every subsystem is built to keep the agent stack under your policy.
/agents/ to add more.ORION is the commander: it takes your objective and keeps execution coherent. Worker agents are the skills — each is a well-defined role with its own tool permissions. The Task Engine runs the plan, retries, and writes state so you can see what actually happened, not a black box.
Drift Guard and circuit breakers are safety rails. Drift Guard notices stalls and silent failures; breakers cut off repeated self-destructive actions (like the same bad install in a loop). Local model layer is where your weights run — default Ollama, or swap the backend if you use another server.
How it works
ORION’s job is to turn intent into a verified result. You don’t stay in the loop for every micro-step, and you don’t lose control of the ones that matter.
Describe what you want in plain language. ORION confirms scope and picks the right model slot for the job.
The plan tree is decomposed into tasks. Each task goes to the worker that declares it, under permissions you set.
Drift Guard watches every heartbeat and re-plans on stalls. You get output with receipts — not hallucinations.
You’re not meant to micro-manage every sub-task. Objective in is where you state intent; Plan & delegate is where work gets assigned to the right agent under your rules. Verified result means the run is observable — heartbeats, logs, and replanning on failure — so the outcome is defensible, not a guess.
When a step needs your approval, it surfaces. When it doesn’t, ORION and the task engine own the loop until the job is done or blocked.
ORION
ORION is the conversational commander. You hand it a goal; it decomposes it into a plan, lines up the right workers and models, drives the task engine with visibility the whole way, and hands back a verified result. Three configurable model slots let you match compute to the job.
Small, fast model for lightweight tasks, acknowledgments, and routing decisions. Typically a 3B–7B local model.
Mid-sized default for day-to-day reasoning and planning. Usually a 7B–14B model that’s comfortable on 16–32 GB unified memory.
The biggest model your hardware can run — 14B, 34B, or a quantized 70B — reserved for complex multi-step objectives.
These slots are not three separate products — they’re where you park models that match the cost/quality of each class of work. A small, fast model is enough for light routing; a big model belongs on the Heavy slot for hard, multi-step reasoning. You can use the same local model in every slot if you want simplicity; the architecture is about flexibility when you care about latency vs. depth.
Size bands (3B–7B, 7B–14B, 14B+ / quantization) are rules of thumb. Unified memory and chip matter: Apple Silicon and enough RAM are what make larger weights practical for local inference.
Trust by design
Privacy at LVLUP isn’t a configurable setting — it’s the architecture. Models, memory, conversation history, agent definitions, and generated output all live on your machine. Nothing phones home unless you explicitly turn an online model on.
Conversations, task history, and memory live in local SQLite under ~/Library/Application Support/LVLUP. API keys for optional online providers are stored in the macOS Keychain via keytar — never written to disk in plaintext.
Capable agents need real tools. Every ORION inference starts with a TOOL-RULES preamble, and dangerous operations (--force installs, edits into non-existent dirs) are blocked outright.
No free tier that sells your context. No analytics, no crash reporting, no telemetry on by default. The deal is the software and the updates — nothing hidden behind the login.
Data stays on-device is the default posture: your prompts, history, and outputs live in local SQLite; optional online providers only get traffic when you wire them in and pick them for a slot. Keys for those providers sit in the Keychain so we’re not leaving secrets on disk in plaintext.
Agents that need to act on your machine get real tools (shell, network, files) with gated, visible rules — the goal is power without “silent exfil” or surprise writes. The $12/month line in the other card is the point: the product’s business is the subscription and updates, not reselling attention on your context.
Pricing
A subscription pays for the software, the update stream, and the right to run it. That’s the entire monetization story — no secondary revenue from your data.
Full access to ORION Control, every worker agent, every future update, and Orion’s Brain in the v1.3.0 public release. One flat price — no tiers, no seat counting, no usage meters for the core product.
You pay US $12/month in Stripe for access to the app and the update path. LVLUP doesn’t monetize your data as a second business — the line above is the core deal.
Your subscription is represented to the app as a short-lived signed JWT. The client caches it so you can work offline for the life of that token, not “phone home” on every message. If a renewal fails (e.g. card), you get a 3-day grace window to fix billing; after that, the app locks until the subscription is good again. Cancel anytime in your account; you keep access through the paid-through date.
No free tier, no refunds (see Terms). US-only at launch — international follows when tax and billing are ready. The fine print in the line above and your Terms cover edge cases; this is the walkthrough.
No free tier. No refunds. US-only at launch; international to follow. Subscription verifies as a short-lived signed JWT so the app keeps working offline. If a renewal fails, a 3-day grace window lets you restore billing before the app locks.
Optional. Billed in Stripe when you use them — separate line items from your app subscription. Nothing here is required to run the product.
Included with your subscription; not a separate billed product.
Per 30-minute block, billed once per block. Integration help, architecture, hands-on sessions — book a block, pay for that block.
Pay in Stripe (30m block) →One-time base; full scope in a quote or SOW before work runs. Quote-scoped, bespoke build.
Pay in Stripe ($997 base) →System requirements
LVLUP Agents AI is a signed & notarized Electron app for macOS. The heavier the model you assign to ORION’s Heavy slot, the more unified memory you’ll want.
Check your system
We use what the browser can report—no app install. Match this against the cards; for a definitive spec use Apple menu → About This Mac on a Mac.
Safari almost never exposes RAM to websites. Chrome and other Chromium browsers may show an approximate GiB value via the Device Memory API; it’s still rounded and can be capped for privacy. Nothing in a normal web page can read your true hardware like the OS can.
These cards are the intended floor for a good experience: a recent macOS, enough RAM for the model sizes you want, and disk for the app + downloaded weights. Apple Silicon is the sweet spot for local LLMs; Intel can still run, especially on smaller models.
The in-page check uses only what a web browser exposes. It’s a sanity check, not a spec sheet: Safari usually hides device memory; Chromium browsers may show a rounded, capped GiB hint. For the last word, use About This Mac on the machine. Nothing here installs software.
Roadmap
v1.3.0 is the paid public release — everything in the first tab ships together. The other tabs are planned direction after that (timelines shift with shipping reality).
Live: v1.3.0 — Clean AI Has Arrived
Local-first AI that plans, executes, and delivers real work — without sending your data anywhere.
Custom auth (bcrypt + JWT + email verification) running on LVLUP’s own backend, backed by DigitalOcean Managed PostgreSQL. Entitlement is a signed JWT cached on-device so offline use keeps working.
Configure provider keys for Anthropic, OpenAI, Google Gemini, Mistral, Groq, xAI (Grok), DeepSeek, OpenRouter, Perplexity, Together AI, and Cohere. Keys live in the macOS Keychain. Every online model is visibly labeled in the UI; the default stays offline.
A top-level tab that pulls down curated known-bugs + fixes, version changelogs, learned session patterns, and structured tool-gap reports — so every LVLUP install gets smarter with each release. Hosted here at ai.lvlup.network.
The dark theme stays the default. A theme-picker architecture on design tokens opens the door for a cream light theme and future skins (Cyberpunk, Retro) without code migrations.
v1.4.0 — “Control & Expansion”
What you’re buying into: I can build my own system.
Purpose: Turn LVLUP from a powerful tool into a system you can shape and extend.
Core theme: “Your AI. Your rules. Your operators.”
Create, modify, and deploy your own agents. Define role, tools, permissions, and outputs. No-code plus advanced config modes.
Turn successful ORION executions into reusable workflows. One-click reruns. Chain multi-step operations.
Granular permissions for file access, shell execution, and network usage. Clear visibility into what ORION can and cannot touch.
Plug in APIs, local scripts, and external services. The tool registry becomes user-visible and controllable.
What this version proves: LVLUP is not just AI — it’s a programmable workforce.
v1.5.0 — “Intelligence That Compounds”
What you’re buying into: My system gets smarter.
Purpose: Move from execution to a learning system.
Core theme: “Your system gets smarter every time you use it.”
Smarter pattern distribution, better task planning over time, shared fixes and faster resolution.
ORION remembers preferences, workflows, and successful patterns. Fully local, user-controlled.
Detect inefficiencies, suggest faster paths, recommend better agents and tools.
Track tasks completed, time saved, and system efficiency. Visual “AI productivity score.”
What this version proves: LVLUP doesn’t just execute — it learns and compounds value.
v1.6.0 — “AI Infrastructure”
What you’re buying into: My system runs my operation.
Purpose: Elevate LVLUP from personal tool to an operating layer for real-world systems.
Core theme: “Run your operation on your own AI.”
Sync multiple LVLUP instances, distribute workloads, local network orchestration.
Multiple users, shared agents and workflows, role-based access.
Background tasks, scheduled missions, persistent agents.
Discover agents, workflows, and tools. Install in one click.
What this version proves: LVLUP isn’t just software anymore — it’s infrastructure.
v1.3.0 (Live tab) is the paid public release — what’s in that tab is the scope we’re talking about for “shipped” in this version. It’s the baseline you buy with your subscription for this major release line.
v1.4+ tabs are forward direction — what we’re building toward, not a dated guarantee. Priorities, sequencing, and scope can change with shipping reality. The timeline strip is an at-a-glance story arc, not a contract calendar.
We keep this page updated as plans firm up. Terms of service and your subscription govern the product in your hands, not a forum post or a single roadmap blurb.
FAQ
Not by default. All models, memory, conversation history, agent definitions, and generated output live in local SQLite under ~/Library/Application Support/LVLUP. Inference runs on localhost via Ollama or a compatible server.
If you enable an online provider (Anthropic, OpenAI, etc.), requests go to that provider under their terms. You choose which slot, which agent, and when.
No. The subscription is $12/month (recurring in Stripe). We don’t run a free tier because the alternatives all involve selling your context somewhere — and that’s the thing we’re rejecting. If you want to evaluate, you subscribe for a month.
The Electron client caches a short-lived signed JWT and re-validates on boot. If your subscription lapses, a 3-day grace window lets you keep using the app with a “subscription lapsed” banner. After that, the app locks until you restore the subscription.
No refunds. The subscription is $12/month and can be cancelled anytime; cancellation stops the next renewal. If you’re running into a technical problem, email us before cancelling — we’d rather fix it.
US-only at launch. International support follows, driven by tax + Stripe configuration work.
In v1.3.0, the app ships with integrations for Anthropic, OpenAI, Google Gemini, Mistral, Groq, xAI (Grok), DeepSeek, OpenRouter, Perplexity, Together AI, and Cohere. Add a key once, and the provider’s models become assignable to ORION’s Quick / Standard / Heavy slots from the same dropdowns you use for local models.
Per-provider monthly caps hard-stop further requests if you blow past a budget.
No. Ollama is the default and easiest path, but the local model layer is an abstraction — any OpenAI-protocol-compatible local server works with a config change. That includes LM Studio, llama.cpp’s server, and vLLM.
The app targets macOS first because Apple Silicon gives us a consistent, capable local-inference platform. A React Native (Expo) mobile companion already ships as a read-only status monitor. Desktop Windows/Linux are not in scope for v1.3.0.
Project LVLUP, LLC — a Wyoming-registered LLC. Support: [email protected].
Download & access
Signed, notarized macOS builds and release notes are published here for subscribed accounts. Create an account, sign in, and the latest build is one click away.