Every feature we ship or change shows up here with plain steps and a button that takes you straight to it. Anything you configure lives in Connect — that's your settings for outside tools.
Connect Nucleus to your AI tools (MCP)
Live
This is how Cursor, Claude Desktop, or ChatGPT-style clients start running every call through Nucleus (estimate · advise · route · meter · readback). One key, paste one config.
Go to Connect → MCP servers.
Click Generate key and copy your nk_live_… key (it's shown once).
Pick your tool — Claude Desktop, Cursor, or Claude Code — and hit Copy on the config.
Paste it into that tool's MCP settings and restart the tool.
Back here, click Test connection — it appears in Live connections. Done: that tool now runs through Nucleus.
Efficiency Planner — spend less, see the whole picture
Live
Before you run a task, see what it costs across models plus power, legal and risk — in plain terms. Nothing to activate; it's advice, you decide.
Open Kitchen → Efficiency.
Pick a task preset and set connectors / turns / data type.
Click Advise — get the cheapest capable model, energy in phone-charges, legal + risk flags, and a Chain readback of how efficiently you actually ran.
What each module is for
Tools — browse every model wrapped in trust, cost, energy, license.
Grade — the one trust/safety score; what's safe to use.
Power — what a run costs in money, energy and carbon (human terms).
Chain — the full record + the readback of how you used things.
Watch — what changed (terms, licenses) and what to do.
Efficiency — plan the cheapest, cleanest way to run a task.
This page updates whenever we add or change a feature — check here first.
✕
Get set up · 2 of 4 · under 5 minutes
Point one app at Nucleus and watch it get graded.
✓
Create workspace
Aluma Studios
✓
Connect a gateway
Cloudflare AI Gateway
3
Connect an MCP tool
One click · you approve first
4
Capture your first recipe
See its trust score
Find the right AI. Use it with confidence.
Every tool arrives wrapped in its trust score, legal profile, energy cost and a full record of where it came from — so you decide with the whole picture.
Tools governed
—
live
Avg trust score
—
from Grade
Energy · your calls
—
from metered usage
Calls logged
—
across your tools
What needs you
Open Efficiency →
Reading your real signals…
Nucleus is checking your usage for anything that needs a decision.
One core · six modules · one energy lens
Model guide
Opus, Sonnet, Haiku, "high effort" — in plain English.
What each model actually does, and which to reach for. No spec sheet required.
Open the guide →
Recent recipes
View all →
Watch · recent activity
Open Watch →
Loading…
your real Nucleus calls will appear here
One line to adopt
Point your app here. Keep your code.
Nucleus speaks the same language as OpenAI & Anthropic. Change one setting and every request is checked — sent to whichever model does it best, wrapped in trust, cost, energy and legal checks, and logged so you can always see exactly what happened.
# connect your AI client to Nucleus — MCP (Streamable-HTTP + JSON-RPC) mcp_endpoint = "https://api.nucleusengine.ai/api/mcp"// estimate + govern headers = { "Authorization": "Bearer nk_live_…" } // your key ↓ MCP servers tab
open — Nucleus works through whatever you plug in
What's a gateway? Think of it as the kitchen's front desk — one place your order goes, which then sends it to the right chef and keeps the receipt. Cloudflare AI Gateway, Ramp Router →
Routing policy
Edit rules →
Default strategyCheapest model that clears the quality bar
Fallback chainRamp Router → AI Gateway → direct provider
GuardrailNever route client-facing renders to research-only models
Monthly budget$250 · $12.40 used
MCP = one standard plug between a model and one tool.like USB-C
Model
can only talk
→
MCP hatch
monitored · you approve
→
One tool
files · chat · database
✓ You approve it✓ Every call logged✓ Close it instantly
Nucleus shows each helper's trust, cost & legal before you plug in — 9,600+ out there.
More detail →
A model on its own has no hands outside the kitchen — it can't open your files, read your calendar, or message your team. The helper advertises a menu of exactly what it can do; the model can only ask for what's on that list; the hatch opens only when you approve it, only for what you allow, and every call is logged so you can watch it and close it instantly. "Standard" is the point — like USB-C, one agreed connection any model can use to reach any helper without custom wiring each time.
What Nucleus Engine gives your AI
6 tools
Once connected, your AI can call these — each maps to a Nucleus module.
AdviseKitchenThe whole picture before you spend — cost, power, legal, risk. Advisory.
EstimatePowerPrices a job before you run it — dollars + energy, in human terms.
RouteGradeRecommends the cheapest model that still clears the trust bar.
GovernChainPolicy + budget check → allow/warn/block, written to the audit.
MeterPowerRecords what a call actually used; burns down your budget.
ReadbackChainAfter you finish: how efficiently you ran + the savings you left.
Your Nucleus key
One secret authenticates every MCP client you connect. Treat it like a password — anyone who has it can spend against your budget. The full key is shown once when you generate it.
# nucleus api key Authorization = Bearer nk_live_…
Add Nucleus to your AI tool
Live connections
Client
Transport
Calls
Last seen
Loading…
Discover more · registry
Local nodes = your own machines (Mac Mini, GPU box) running models. Point Nucleus at them so runs cost ~$0 and never leave your network — the greenest, cheapest option when a local model can do the job.
Tool
Trust
Security
Performance
Transparency
Status
Live from Grade — every tool your AI touches is scored on security, performance and transparency, blended into one trust score. Full cost + energy per tool lands with the inventory endpoint.
Nucleus Kitchen · Use AI With Understanding
Nucleus Kitchen is where AI stops being a mystery.
Kitchen is the whole space around the AI — not one thing. It explains every model in plain language, shows you what's safe to use and what it'll cost, lets you actually cook (run the work), and saves what worked so you can do it again. A recipe is just one part of the kitchen, not the whole thing.
What the kitchen gives you · tap one
Understandevery model, in plain words
Choosethe right tool, safely
Cookrun the work
Reuserecipes, skills & prompts
Tap a part of the kitchen to see what it means →
Recipes, skills & “.md” files · your cookbook
The written cards the chef follows — so you get the same result every time.
These three words scare people off. They shouldn't — they're all just paper. Tap one.
A recipea job you saved
A skilla rule you wrote
A “.md” filethe paper it's on
Tap one to see what it really is →
Put simply: a recipe is a job you saved. A skill is a written instruction you hand the AI so it works your way. A “.md” file is just the plain-text page a skill is written on — like a handwritten recipe card. None of it is code — it's all things you can read and write yourself.
Recipes · captured workflows
+ New recipe
Back to Kitchen
Recipe · render pipeline
Shadows in the Light
Workflow — 6 tools
directed graph
Grade · trust
72
/ 100 · workflow
Reproducibility91
Legality confidence44
Cost efficiency88
Legal profile
Not cleared for commercial
Flagged toolStable Video Diffusion
LicenseStability Community · non-commercial
This workflowpaid client deliverable
Data handlinglocal — never leaves node
Fix: swap SVD for LTX-Video (permissive), or add a Stability commercial license.
local · background graded, hero untouched · R2 · 04:48:40
Kitchen · Efficiency & Tool-Chaining
How efficiently your AI is actually running.
This reads back your real usage — every tool call your AI made through Nucleus — and tells you cost, energy, and where you overspent, with the reason why. Nucleus advises; you decide. Live from api.nucleusengine.ai.
Your usage — the real picture
reading…
Reading your real usage from Chain…
⟐ What-if planner — estimate a task before you run it (hypothetical, not your real usage)
Sketch a task and Nucleus estimates the whole picture across models. This is a planning tool — the numbers above are your actual usage.
Model price book · live · Aug 2026
Advisory only — Nucleus shows the whole picture; you decide and enforce. Power figures shown in human terms; legal (Watch) + risk (Grade) are evidence-backed in the live modules.
Workspace & account
WorkspaceAluma Studios
Signed inJP · jerome@alumastudios.com
PlanNucleus · Info@ account
Keys & connected tools
Your Nucleus key and the MCP install steps for each tool live in Connect.
Defaults
Advisory defaults for the Efficiency lens. Nucleus advises; you decide — these never block anything.
Appearance
ThemeDark (default)
DensityComfortable
Light theme + density options are on the roadmap.
Advanced (developer)
MCP endpointhttps://api.nucleusengine.ai/api/mcp
API basehttps://api.nucleusengine.ai
Anonymous MCPallowed (MCP_ALLOW_ANON=1)
Overview
Kitchen · Learn to use AI with confidence
From "fancy Google" to genuinely good at this.
Most people meet AI as a chat box that sort-of does things. This takes you — one comfortable step at a time — from what it actually is, to picking the right tool, to using it efficiently. Start at level 1. Go as far as you like.
At a glanceStep 1 of 8
The basics · mostly pictures
How it works, at a glance.
The whole thing, as a kitchen
The AI is the chef. Everything else is the kitchen around it.
The modelthe chef
Tokensyour words, diced
Contextthe counter
Effortcook time
Reaching outa safe door to one helper
How a model is made
How a chef becomes a chef.
A model isn't built like an app — it's trained like a cook. Tap a step.
1Reads everythinga huge library
2Practises endlesslyguess & correct
3Gets feedbacka tasting panel
4Becomes a modelready to cook
Tap a step to see what happens →
What the model is actually doing
01
It predicts the next word
Over and over — that's the whole trick.
02
Billions of tiny dials
More dials = smarter, but slower & pricier.
03
Words become pieces
You pay per piece — in and out.
tokens · the thing everyone keeps talking about
Type anything below — watch the AI chop it into tokens, live:
The strawberry sat on the windowsill.
Tokens0
≈ words0
≈ cost to read$0
Memory used0%
of a small 8,000-token “desk”
Tokens: how the AI actually reads and writes — not letters, not whole words
Your words and files are the ingredients; the AI dices them into countable pieces — tokens — and answers back in tokens too. Rough ruler: 1 token ≈ ¾ of a word.
Cost — billed per token, in and out.
Memory — holds only so many at once (its “desk”).
Speed — one at a time, so long answers take longer.
Also why “how many R’s in strawberry?” trips it up — it only saw “straw” + “berry,” never the letters.
why it can’t spell
Why it miscounts letters
It sees two chunks, never s-t-r-a-w-b-e-r-r-y — so counting the R's is guesswork.
a rough ruler
A rough ruler
Handy for guessing size and cost before you run something big.
04
A counter, not a memory
Wiped clean when you close the chat.
05
Frozen at a date
No live news — unless you give it search.
06
Thinking = taking its time
More time, better on hard tasks.
What it knows, its energy & where it runs
07
It burns real electricity
Measured on your own machine.
08
Rent it, or own it
Cloud (easy) or your machine (private).
09
Where cloud AI lives
Giant warehouses of computers — "data centers" — that gulp power. That's what you rent.
10 · two kinds of chip — AI runs on the GPU
AI = a million tiny identical sums, all at once
Two kinds of chip — and why AI runs on the GPU, not the CPU
The CPU is your computer's everyday brain — one job at a time. The GPU is a whole brigade — thousands at once. AI is millions of tiny identical sums, so it runs on the GPU. (That parallel muscle is also why GPUs guzzle power.)
try it · the same job on both chips
CPU one at a time
0 / 12
GPU all at once
0 / 12
Same 12 little sums. Press play and watch how each chip gets through them.
See it: one at a time vs all at once
Give both chips the same twelve sums. The CPU works through them one after another — twelve steps. The GPU does all twelve at the same time — one step. Now picture millions of sums: that's every AI answer, which is exactly why it runs on the GPU.
Pictures, video & reaching your stuff
11 · pictures & video
your words guide it
How it "cooks" a picture — and why video costs so much more
It never paints pixel by pixel. It starts from a full screen of TV-static, then in many small passes it wipes away a little of the fog each time — every pass nudged toward what your words asked for — until a clear picture appears. The name for this is diffusion: static → rough shapes → finished photo. A video is dozens of these pictures every second, and the hard part is keeping them consistent — the same face, the same light, smooth motion from one frame to the next. All that extra "stay consistent" work is why video is far slower and pricier than a single image (your Stable Video Diffusion (SVD), Flux and LTX-Video models).
12 · Model Context Protocol (MCP)
A Model Context Protocol (MCP): the universal plug between the AI and your tools
the helper's menu →read a filelist your eventssend a draft
On its own the model is a brilliant cook sealed in the kitchen — powerful, but cut off: no deliveries, no phone, no way to reach your files or calendar. An MCP changes that. Each helper arrives with its own menu, written in plain language, saying exactly what it offers — “read a file”, “list your events”, “send a draft”. The model just reads the menu and places an order; the result comes back. It never needs to know how the helper works inside (you don't need to know how a restaurant's kitchen runs to order off the menu) — and nothing happens without your approval.
Why everyone's talking about it: before MCP, connecting each tool to each AI was a custom wiring job — a different cable for every pair. MCP makes it one standard plug — think USB-C, or an app store where every app comes with its own manual — so any AI can discover and use any MCP tool. It's the moment AI went from smart but isolated to actually able to reach out and act in the world.
The one-minute version
It's not a search engine. It's more like a very well-read assistant.
Google hands you links to pages. AI writes a fresh answer in its own words — like asking a sharp colleague, not searching a library. Great for drafting, summarising, translating, brainstorming.
The catch: it can sound sure and still be wrong — it writes what sounds right. Fix: give it the facts (paste your doc, or let it search).
You stay in charge. It drafts and suggests — you decide. That's the whole idea: it amplifies you, it doesn't replace your judgement.
You don't need the names yet
There are different AIs for different jobs. Just tell it how hard the task is.
Slide to match your task
Simple errand, or something you really need to get right?
The only three words worth knowing
"Model"
Just means one AI. There's a family of them — some small and quick, some large and thoughtful. Same idea, different sizes.
"Effort" / "thinking"
How long it takes to answer. Quick for easy things; slower and more careful when it needs to work a problem out.
"Context"
How much it can look at right now — a paragraph, or a whole document you give it. It's a desk, not a memory: it forgets once you close the chat.
That's genuinely enough to start. When you're curious how it actually works — or what all the money-and-power talk is about — level up. No pressure, no rush.
The quickest way to get it is to watch it happen. Four tiny live examples — nothing to memorise.
1 · How it answers you
it guesses the next word, over and over
Under the hood it does one thing: look at the words so far and pick the most likely next one. Watch it build a sentence a word at a time — that's the whole trick, repeated fast.
2 · "Thinking" is just taking its time
same question, same AI
Ask "what's 17 × 24?" two ways. On the left it blurts a quick guess. On the right it writes out the working first — slower, but right. That's all "more effort" means.
Quick answerfast & cheap
Takes its timeslower · pricier
3 · What a Model Context Protocol (MCP) is
giving the AI one helper it can reach
On its own, an AI can only talk — it can't reach your files, your email, or your calendar. An MCP is a safe doorway to one outside helper. It only opens when you say yes, and only for the one thing you allow. Watch a request go out through the door and come back with just what you approved.
The modelcan only talk
you approve
A helperyour files
4 · Why longer = more expensive
it charges by the piece
Your words get chopped into little pieces called "tokens" (about ¾ of a word each). You pay for every piece going in and coming out — so a longer question, or a longer answer, simply costs more.
Cost
$0.0000
Compare the models side by side — capability, speed and cost — and pick the right one for the job. This is where you become the person who chooses well and spends wisely.
Best for
Skip for
What the four dials actually mean
Every model trades these four off against each other.
There’s no single “best” model — only the right one for the job. These are the levers you weigh.
Reasoning
How well it works through hard, multi-step problems — logic, tricky code, careful analysis. Higher means fewer slips on the difficult stuff.
In the kitchen: the cook’s skill on complicated dishes. A short-order cook nails eggs; a master pulls off a 12-course tasting menu.
More reasoning is slower and pricier — worth it only when the task is genuinely hard.
Speed
How fast the answer comes. It writes one token at a time, so this is really “tokens per second” plus how quickly it starts.
In the kitchen: how fast plates come out of the pass.
Smaller models are usually faster — great for quick, simple jobs.
Cost
What you pay, counted per token in and out. Bigger models charge more per token.
In the kitchen: the price of the ingredients plus the chef’s time for each dish.
Match the model to the job — don’t pay head-chef rates to boil an egg.
Context
How much it can hold in front of it at once — your question plus any documents you give it — measured in tokens. When it fills up, the oldest bits fall off the edge.
In the kitchen: the size of the prep counter: a big one fits a whole cookbook, a small one just a recipe card.
Big context reads long documents, but filling it costs more and can slow things down.
Other words you’ll hear — and what they really mean
No, the four dials aren’t the whole story.
These get thrown around a lot. None of them are as scary as they sound.
Reasoning / “thinking”
Newer models can think before they answer: they write out a private chain of steps — a scratchpad — catching and fixing their own mistakes along the way, then give the final answer. It's deliberately trained in, and it makes hard problems (maths, logic, multi-step code) far more reliable. But more thinking = more tokens = slower and pricier, so it's saved for the genuinely hard stuff. It's still next-word prediction — just given room to work.
In the kitchen: a line cook plates the first thing that comes to mind; a head chef tastes and adjusts through every step before the plate leaves the pass. Same chef — just taking the time to get it right.
Weights / parameters
The millions-to-billions of tiny numbers the model tuned while learning. They are what it knows — its instincts. “A bigger model” mostly means more of these.
In the kitchen: the cook’s muscle memory and seasoned instincts, built over years. More of it = a more capable, but heavier, slower, pricier cook.
Training vs. inference
Training is the one-time, months-long, hugely expensive learning — done once, by the maker. Inference is the model running to answer you; every reply you get is one “inference.”
In the kitchen: training = the years spent becoming a chef. Inference = cooking your specific order, right now.
Knowledge cutoff
The date its training stops. It knows nothing that happened after — unless you give it web search or paste it in.
In the kitchen: the last cookbook it read. No newer recipes unless you hand them over.
Multimodal
Whether it handles more than text — images, audio, sometimes video — reading them, making them, or both.
In the kitchen: a cook who doesn’t just cook but can also plate, photograph and taste — works in more than one medium.
Temperature
A dial for how adventurous vs. focused the answers are. Low = safe and repeatable; high = more varied and creative.
In the kitchen: how much the cook improvises versus following the recipe to the letter.
Fine-tuning
Extra, focused training added on top of a general model to make it excellent at one specific thing.
In the kitchen: a trained cook who then did a specialist pastry course.
AI Explained · social edition
Get AI, one swipe at a time.
No essays. Move through it like a story — tap the right for next, left for back (or drag, or use arrow keys).
nucleus.kitchen· AI, explained
340swipe →
AI & Coding · in plain English
Yes, it writes code. Here's the trick.
Code scares people off. It shouldn't — it's just instructions, and the AI writes them the same way it writes a sentence.
The whole thing · tap one
Codeprecise instructions
How AI codespredicts the next piece
Vibe codingdescribe it, don't type it
The toolsa co-pilot in the editor
Your jobtest & approve
Tap one to see what it means →
watch it · autocomplete
It finishes your thought — in code
You write a plain-English note; it predicts the code that should follow, one piece at a time. Exactly the same next-piece trick as finishing a sentence.
01
Code = a recipe for a machine
Exact instructions, written so precisely a computer follows them with zero guessing.
02
It writes code like it writes words
One piece at a time, predicting what comes next — learned from millions of real programs.
03
“Vibe coding”
Describe it in plain English → it writes the code → you run and refine. You never touch the syntax.
04
Brilliant — but not always right
Great at boilerplate, translating, explaining and fixing. Slips on subtle logic and security. So: test everything.
In Nucleus, a bit of code that works becomes a reusable recipe — scored for trust, with its cost logged — so good solutions get captured, not lost.
AI & Coding · swipe edition
Coding, one swipe at a time.
The short, swipeable version — tap the right for next, left for back.
nucleus.kitchen· coding
210swipe →
Your AI · Inventory
Every model your AI actually uses — in one place.
Nucleus watches which models your tools reach for, then builds your own catalog: what each one is, what it costs, how hard you lean on it, and how much energy it burns. This is your AI, measured — not a demo.
Models you're using
reading…
Reading your usage from Chain…
Where your AI connects
—
Built from your real Chain records + the live model price book. A model appears only once your AI has actually used it through Nucleus — nothing here is seeded.
Nucleus Grade · Trust & Safety
Nucleus Grade is your food-safety inspector.
Before a dish reaches a guest, someone checks it's safe, honest and allowed. Grade does that for every model and recipe — then turns it into one simple rating.
The inspector's checklist · tap one
Reliable?works every time
Safe?no nasty surprises
Allowed?licence & rules
Private?your data, safe
Tap a check to see what it means →
It all becomes one rating — the trust score
78
out of 100
Reliable88
Safe82
Allowed (legal)61
Private90
One number, 0–100. High enough → safe to use. Too low → Nucleus can quietly hold it back until it's fixed. Like a hygiene rating on the door.
Nucleus Power · Energy & Cost
Nucleus Power is the meter on your stove.
Every dish burns something — time, money and electricity. Power shows what each run costs, before you cook and after — so nothing is a surprise.
Show energy as
Spend · your metered calls
—
from real usage
Energy · your calls
—
metered through Nucleus
Carbon
—
CO₂e equivalent
What the meter tracks · tap one
Moneywhat it costs
Energyreal watts
Carbonits impact
Before vs afterestimate → truth
Tap to see what the meter tracks →
Wait — what's a "watt-hour"?
A watt is how big the flame is — how fast electricity is being used. Leave it running for an hour and you've used that many watt-hours (Wh). It's the exact unit on your home electricity bill — 1,000 Wh = 1 kWh, the number the power company charges you for.
To feel the size
~15 Whcharge your phone once
~10 Wha lightbulb for an hour
~100 Whboil a kettle of tea
~250 Wha laptop all day
~1,000 Whyour fridge for a day
So the "Shadows" render at ~118 Wh ≈ boiling one kettle of tea. Tiny on its own — but across thousands of runs it adds up, which is the whole reason Power keeps count.
How the numbers are worked out
plain arithmetic
Energy
flame size watts×time on hours=energy Wh
Exactly like a stove: a bigger flame, on for longer, uses more. On your own machine we read the real watts off the GPU while it cooks — so it's measured, not guessed.
Money
pieces tokens×price per piece=your bill $
Cloud models charge per token (see it move in Level 2 of AI explained). Cook on your own stove and this is basically $0 — you already own the kitchen.
Carbon
energy Wh×how dirty your grid is=carbon CO₂
The very same dish is greener on a clean grid. We use your region's power mix to work it out.
Worked example — "Shadows in the Light": the render drew about 104 Wh measured on the NVIDIA card (a real meter reading), plus ~2 Wh estimated for the caption step in the cloud → about 118 Wh and ~47 g CO₂ for the whole clip. Local reads real; cloud is estimated from tokens and the provider's published rates.
Nucleus Chain · The Paper Trail
Nucleus Chain is the order ticket for every dish.
In a good kitchen, every plate has a ticket: what went in, which station touched it, when. Chain keeps that for every AI result — so you can trace any output back to exactly how it was made.
What's on the ticket · tap one
What went inthe ingredients
Which stationsevery tool, in order
Whenthe kitchen clock
Why it mattersproof, on demand
Tap to see what's on the ticket →
A real ticket · "Shadows in the Light"
Caption model → wrote the motion prompt
in: hero.jpg · out: prompt · 04:44:07
rembg → cut out the subject
local · hero.png · 04:44:31
SVD → made the base video
florida-gpu · svd_xt · 04:47:52
ffmpeg → composited & delivered
hero untouched · R2 · mp4 · 04:48:40
Nucleus Watch · Keeping an Eye Out
Nucleus Watch is the head chef running the line.
A good expediter watches every station, calls out a problem before it reaches a guest, and notices the moment a supplier changes something. Watch does that for your AI — quietly, in real time.
What the head chef does · tap one
Watches the linereal-time
Calls out problemsbefore you serve
Notices changeswhen suppliers shift
Keeps it safeauto-pause & swap
Tap to see what the head chef does →
Live activity — your real calls
—
Loading your activity…
every Nucleus tool call your AI makes shows up here
Nucleus Vault · The Locked Cabinet
Nucleus Vault is the locked cabinet in the kitchen.
Every kitchen has a locked cabinet for the valuable stuff — the keys, the safe, the good knives. Vault holds your passwords and access keys so tools can be used, but never handed the keys directly.
How the cabinet works · tap one
Keys & passwordsthe valuables
Locked awaynever shown
Used, not handed overthe tool never keeps it
House ruleswho & how much
Tap to see how the cabinet works →
In the cabinet
names only — values stay hidden
Anthropic keysk-ant-••••••••
OpenAI keysk-••••••••
Cloudflare token••••••••
OpenWebUI key (local)••••••••
You'll only ever see the name of a secret here — never the value. That's the point: even the cabinet's own screen keeps it hidden.