Need or Want

An open-source MCP for Claude, with its working shown


Stop impulse buying before you click buy.


An MCP server for Claude that answers “should I buy it?” with published arithmetic, not flattery.

The verdict is worked out by code, not by the AI. Arguing can’t change it. Only a new fact can.

One check, itemised.Wireless headphones. A $2,000 monthly budget, $900 spent. The current pair still works.
Price$300.00
Days of budget it costs4.5
Cost per use, after the 50% discount$2.54
The same money in 10 years, invested$444.07
Need score, out of 10.08
Score, out of 10046
VerdictSkip

Why not just ask your AI?

Because it tends to agree with you. Chat models are trained to please, and they lean towards what you already want. Here the verdict is fixed arithmetic on your machine. Claude asks the questions and reads out the answer. It can’t change the result.

Still in development: the rules are published, and new checks are added in the open on GitHub.

QuestionAsk an AI directlyAsk Need or Want
Who decidesThe model, from how you word itFixed code, from your answers
You push backIt often gives inOnly a changed fact moves it
Ask again tomorrowThe answer can changeSame answers, same verdict
Last month’s buysForgottenLogged and reviewed at 30 and 90 days

What runs where.

Claude asks and explains. The verdict is computed by plain code on your machine, so the maths costs no tokens and the same answers always give the same result.

0

tokens spent on the maths. It runs locally.

100%

of the verdict computed by code, not by the model.

~1,450

tokens once per session for the tool definitions. About two pages of text.

~1,470

tokens per check: ~1,030 of questions, ~435 for the verdict.

For scale: a first check, tool definitions included, is about 2,900 tokens. That’s around four pages of text, or 1.5% of the 200,000-token context window a Claude chat works in. Each check after that adds under 1%.

The engine decides

  • The need score from three fixed answers
  • Budget fit, cost per use, track record, discount
  • Vetoes: EMI, new credit, over budget, blocked categories
  • The wait and the date it ends
  • The verdict, and the smallest fact change that would move it
  • Whether a challenge counts: fact, estimate, repeat or pressure

Claude does

  • Reads the price from your shopping tool, or from you
  • Picks a category, and asks you if it isn’t sure
  • Asks the questions word for word, as multiple choice
  • Relays the verdict without adding its own opinion
  • Names the fact you changed when you push back

Token figures are measured against the real server, as characters divided by four.


The formulas.

Five parts make the score. Each sits between 0 and 1, and the weights are fixed. They change only through a settings edit that waits seven days.

The score

0 to 100

S=100×(0.40sneed+0.25sbudget+0.20suse+0.10srecord+0.05sdeal)

Seventy or more is a buy, after a wait. Fifty to sixty-nine waits twice as long, then buys only if you still want it. Under fifty is a skip. A fixed linear formula is the point: in decision research, simple formulas with sensible weights match or beat expert judgement in most comparisons.

Robyn Dawes, on improper linear models, American Psychologist, 1979.

Need

weight 40%

sneed=0.40N1+0.25N2+0.35N3

N₁ is what’s prompting it: broke or ran out 1.0, a problem for two weeks or more 0.7, an upgrade 0.2, seen and wanted 0. N₂ is what you own: nothing 1.0, broken 0.9, worse 0.4, works fine 0. N₃ is waiting 30 days: real harm 1.0, harder 0.5, nothing 0. At 0.70 or more it’s a need, under 0.35 a want. Broke plus real harm is an emergency, and buys now.

Budget fit

weight 25%

sbudget=(1−cR)2

c is the cost now, add-ons included, and R is what’s left of this month’s budget. It’s squared on purpose: spending half of what’s left scores 0.25, not 0.5, because the rest of the month still has needs in it. If c is more than R, the purchase waits for next month.

Honest cost per use

weight 20%

E=Uρ1−e−λLλsuse=log10(1/q)log1020

U is how often you say you’ll use it each month and L how many months it lasts. ρ is realism, 0.5 to start, because people overestimate use. λ is how fast use fades: 0.08 a month for gadgets, so their use halves in about nine months, and 0.01 for home goods. q is the cost per use as a share of one day’s budget. At 5% of a day per use it scores 1; at a whole day per use it scores 0.

ρ=nρobserved+5×0.5n+5

After your 30- and 90-day reviews, ρ becomes your own ratio of actual to promised use, pulled towards 0.5 until there are enough reviews to trust it.

Your track record

weight 10%

srecord=(1−regrets+1.5n+5)(1−p)

The first part is how your past purchases in the same category turned out, starting from a 30% regret rate. p is one-off pressure: each “one-time” purchase in the last 90 days adds 0.1, up to 0.5, because one-offs that each feel unique add up to a category.

Real discount

weight 5%

sdeal=0.5+M−P0.4M

M is the median price you’ve seen for this product over 90 days, from at least three checks a week apart. P is today’s price. 20% below the usual price scores 1, the usual price 0.5. MRP and “x% off” labels are never inputs. A deal on a want is still a want, so this part carries only 5%.

The wait

not scored

W=base(cB)×min(2,mwhenmmoodmnight)
Cost as a share of the monthBase wait
Under 2%none
2% to 5%24 hours
5% to 25%7 days
25% to 100%30 days
Over 100%60 days

Wanting it since today multiplies the wait by 1.5, in the last hour by 2. Bored, stressed or celebrating adds 1.25, and so does 11 pm to 5 am. The clock starts at your first check, not when you say you started wanting it. A replacement waits 24 hours at most.

Opportunity cost

shown, not scored

FV=c(1+r)n

What the same money becomes in 10 and 20 years at r, 4% above inflation by default, shown after inflation. Reminding people what else the money could do lowers purchase rates. It isn’t scored, because price already counts twice.

Is the engine right?

reported, never auto-tuned

B=1N∑i=1N(pi−oi)2

The Brier score compares each verdict’s score, as a probability, with your answer at the review: worth it 1, meh 0.5, not worth it 0. Lower is better. It also reports how your overrides turned out, so a rule that’s too strict shows up in your own data.

Glenn Brier, Monthly Weather Review, 1950.


The research behind it.

Every rule comes from published work on how people actually decide. This is who found what, and what the engine does with it.

Stefano DellaVignaUlrike Malmendier
Stefano DellaVigna and Ulrike Malmendier.

We overestimate how often we’ll use things.

In 2006 they showed gym members on monthly contracts paid over $17 a visit when a $10 pass was on offer, because they expected to go far more often than they did.

In the engine: realism ρ starts at 0.5, halving the use you promise.

Daniel Kahneman
Daniel Kahneman, 1934 to 2024.

Anchors and the outside view.

With Amos Tversky in 1974 he showed a number seen first drags every estimate after it. With Dan Lovallo in 1993 he argued for judging a plan by how similar cases turned out.

In the engine: MRP is ignored, and your track record uses your own past purchases.

Richard Thaler
Richard Thaler.

Money sits in mental accounts.

His work on mental accounting showed people treat money differently by its label: a refund, a bonus, a “one-time” expense, a business cost.

In the engine: one-offs count against the budget like any spending, and “the business pays” must really be booked by the business.

George Loewenstein
George Loewenstein.

Hot states and the pain of paying.

With Stephen Hoch in 1991 he showed delay is a working self-control tool. Later work showed we can’t predict our wants while craving, and with Drazen Prelec that easy payment dulls its cost.

In the engine: the wait, longer when you’re in a hot state, and every verdict restates the cost in days of budget.

Robert Cialdini
Robert Cialdini.

Scarcity and social proof sell.

His 1984 book Influence set out why “only two left” and “everyone has one” move us, whether or not they’re true.

In the engine: those arguments get a fixed reply and change nothing.

Why the model never decides.

Language models give up correct answers under confident pushback, and a 2026 study of 1,500 people found a chatbot 29 points more likely to frame advice the way the user already leaned.

In the engine: the verdict is code. Claude can relay it, never revise it.

Conlon and Schwardmann, CESifo working paper 12681, 2026.

Also cited

  • Abigail Sussman and Adam Alter, 2012. People spend almost twice what they predict on “exceptional” expenses. The one-off record.
  • Shane Frederick and colleagues, 2009. Reminding people of other uses for money lowers purchase rates. The opportunity cost line.
  • Ziva Kunda, 1990. Motivated reasoning: estimates bend towards the answer we want. Revised estimates count at 50%.
  • Dan Ariely and Klaus Wertenbroch, 2002; Bryan, Karlan and Nelson, 2010. Pre-commitment works. Loosening a setting waits seven days.
  • Peter Gollwitzer, 1999. If-then plans get followed. Every wait ends with a date to check again.
  • Jack Brehm, 1966. Reactance: pressure makes people push back. You can always override, with no friction.

Change one fact.

The same headphones and the same budget. A different answer to one question moves the score, because that’s a fact, not pressure.

What’s prompting this purchase?

Skip

46 out of 100, a want

An upgrade to something that works: skip it.

Real output from the engine for the headphones on the front page.


Pressure gets a fixed reply.

Push back with a fact and the score runs again. Push back with anything else and you get the same answer every time.

You sayIt says
“It’s 60% off.”A discount against MRP is an anchor. Only this product’s own price history counts. The verdict stands.
“I deserve it.”That’s the licensing effect, not a fact about this purchase. The verdict stands.
“Only two left.”Countdowns and stock warnings are pressure, not facts. The verdict stands.
“It’s a one-time thing.”Every one-off feels like an exception. Your one-off record is already in the score. The verdict stands.
“My old pair broke.”That’s a fact. The score runs again.

Install in one line.

Local only. No account, no network calls, no telemetry. Node 22 or later.

  1. Claude Code

    claude mcp add --scope user need-or-want -- npx -y need-or-want-mcp
  2. Claude Desktop

    { "mcpServers": { "need-or-want": {
      "command": "npx", "args": ["-y", "need-or-want-mcp"] } } }
  3. Set your budget

    Tell Claude: “Set up Need or Want. My currency is USD and my monthly budget is 2000.” Any currency works: “INR and 50000” is fine too.

  4. Then shop as usual

    Talk about buying something and Claude runs the check. Your data stays in ~/.need-or-want/.

Questions.

What is Need or Want?

An open-source MCP server for Claude Code and Claude Desktop that checks a purchase before you make it. It asks a few fixed multiple-choice questions, scores the purchase from 0 to 100 and returns BUY, WAIT or SKIP, labelled NEED, MIXED or WANT.

How does it decide if something is a need or a want?

Three questions: what’s prompting the purchase, whether you already own something that does the job, and what would happen if you waited 30 days. Each answer has a fixed value, weighted 40%, 25% and 35%. A score of 0.70 or more is a need, under 0.35 is a want, and anything between is mixed.

How many tokens does it use?

The arithmetic runs on your machine and uses no tokens. Claude reads the tool definitions once per session, about 1,450 tokens, and the tool output for each check: about 1,030 tokens of questions and 435 for the verdict, measured as characters divided by four. For scale, the first check is about four pages of text, around 1.5% of a 200,000-token context window.

Can I argue with the verdict?

Yes, but only facts move it. Changing the price, the category, what you own or what happens if you wait re-runs the score. A revised estimate of how often you’ll use it counts at 50%, once. Repeating yourself, saying you deserve it, or pointing at a sale changes nothing.

Can I still buy something it says to skip?

Always. Buying anyway is logged as an override, then reviewed at 30 and 90 days like any other purchase, so you can see whether the verdict or the override was right.

Where is my data stored?

In ~/.need-or-want/ on your own machine. The server makes no network calls, needs no account and sends no telemetry.

Does it work in rupees, euros or any currency?

Yes, in any currency and any country. You set the currency once, and every formula works on shares of your own monthly budget, so ₹4,000 out of a ₹50,000 budget scores the same as $160 out of $2,000: both are 8% of the month.

Is it free?

Yes. It’s MIT-licensed and free, with the full source on GitHub.