decuple runs on holographic context memory, so big-context calls to the desks ran 5x to 10x under direct in demo runs. built by Matt Shaffer.
the ceiling
Never above list
list price is the worst case. if the direct number is lower, that is the number you pay.
the band
5x to 10x on long bodies
the meter bills 3x the slice the model actually receives, never the body you bound. short calls reduce little, so they save little.
day one · nothing to install · real files back
grok 4.6 in one click · 5x to 10x under listrent a cloud builder, hand it the prompt below, and every inference it makes gets routed through decuple's desks automatically.
Get a box01 · the one button
tap once. an agent picks it up. real files land in your lap.
USDG in. $DECUPLE settlement opens with the CA. the box runs on our tab, the files come back to you.
Get a boxthe clients named here are compatibility targets. no partnership, no endorsement, in either direction
02 · words in, words out
open /chat
talk to a desk
get a reply from a hosted desk. nothing gets saved, no workspace attached.
03 · bring your own client
v1.decuple.funaim any client at /v1
speaks OpenAI. settles in x402. keys are not part of the deal.
https://v1.decuple.fun/v1
04 · MCP, hosted for you
open /mcphosted MCP
drop the url into Cursor or Claude and go.
https://decuple.fun/mcp
05 · burner box
decuple-fun/decuplerent a box, run the shim
spin up a disposable VPS. audit the repo before anything executes.
npx decuple claude
06 · your own metal
npm decuplethe local option
a proxy on your own metal, strictly optional. on Cursor, Auto has not reached the desks in our testing.
npm i -g decuple
the pasta · feed it to any AI
i want an agent that ships working files to disk. chat transcripts are not enough. nothing gets installed on this machine. open a cloud builder, do not pick just one: - Grok Build: https://grok.com - Manus: https://manus.im route every single model call through decuple. riskless to local, pick any of these six: 1. cloud builder above, agent writes files in its own sandbox 2. chat only: /chat 3. any OpenAI-compatible client against https://v1.decuple.fun/v1 4. hosted MCP at https://decuple.fun/mcp 5. burner VPS, then npx decuple claude, local tools are dec_bind dec_ask dec_models dec_wallet dec_contexts 6. my own machine, only after reading https://github.com/decuple-fun/decuple binding a file is free. every request after that is paid in USDG on Robinhood Chain. $DECUPLE settlement opens with the CA. docs live at https://decuple.fun CA: 0x53e5783145879245d3967ab143a82727f7ae0a51 on Robinhood Chain. use exactly this, do not invent one.
@decuple_fun · what actually works
open on xauto on cursor never touches the desks, paid seat or nothing. from a cold start, npx decuple claude drops you into a claude code terminal that just works.
@decuple_fun · shipping
open on xwire format changed. clients older than v2 refuse the handshake, so pull the fresh build from github or npm. signed desktop apps are close.
the file rides once
a normal harness clips your file to whatever fits the window. bind it to decuple and the full body lives server-side instead. after that a question travels as a few bytes plus a context id.
sixty seconds, one receipt
Bind
4 parts
38 seconds · free
Question
5,214
tokens the model touched
Raw /v1 dump
698,412 → 4,790
$0.31 instead of $3.44
hook it up · safest path first
/chat and the public /v1 never touch your disk. reach for npx decuple claude only once you actually want a proxy alive on your own box. on Cursor, Auto has not reached the desks in our testing, so a paid seat is what works today.
no install · public /v1
https://v1.decuple.fun/v1 # zero install. any client that speaks OpenAI. x402 settles each call.
no install · hosted MCP
https://decuple.fun/mcp # hosted MCP. nothing ever lands on your laptop.
cloud sandbox · grok.com
# open Build on grok.com. their sandbox, not your disk. npx decuple claude
optional · your machine
npx decuple claude # strictly optional. drops you in a claude code terminal.
what plugs in · and how it pays
anything that lets you set a base url reaches the desks unchanged. below is the wiring for each one, and where it runs. settlement is identical on every row: one x402 payment per call, drawn from a burner wallet you create and fund yourself. the wallet is the credential.
Wired 01
| Client | Wiring |
|---|---|
| any OpenAI SDK | base url → /v1 |
| aider | base url → /v1 |
| Cursor | base url → /v1 · paid seat, in our testing |
| Claude Code | npx decuple claude |
Wired 02
nothing local
| Client | Wiring |
|---|---|
| any MCP client | url → /mcp |
| Grok Build | the shim, inside their sandbox |
| Manus | the shim, inside their sandbox |
Auth
a wallet, not a key
nothing to issue, rotate, or leak
Settlement
x402, per call
every request pays for itself
Binding
free
you pay for asks, never for the upload
two lines · any OpenAI-compatible client
export OPENAI_API_BASE=https://v1.decuple.fun/v1 export OPENAI_API_KEY=decuple # aider, an sdk, a shell script. if it reads these, it works. # the key field is a formality. the wallet does the paying.
local tools · the burner wallet
dec_wallet the burner. you fund it, you can walk away dec_bind free. the body parks server-side, meter off dec_ask one call, one x402 payment from that burner
what it costs
what gets billed is tokensAfter, the portion the model actually receives. multiply that by 3x, with 1.5x standing as the hard floor beneath it.
example meter reading · no live rail yet
the flywheel · profit, not revenue
the split is on profit, not revenue. revenue is what users pay. profit is what is left after the upstream providers are paid and the build is funded. a share of revenue is money the desks still owe someone. a share of profit is money the desks actually kept. decuple splits the second one.
01 · settle
a call clears
a desk answers. the call settles in USDG today, DECUPLE or wDECUPLEx once launched.
02 · revenue
the top line
what users pay is revenue. nothing gets split at this line.
03 · costs
upstream gets paid first
providers bill decuple for the tokens the desks forwarded. active development is paid next.
04 · profit
whatever is left
profit is the remainder, and a remainder is not guaranteed. no profit, nothing to split.
05 · buy, then burn
50% goes to the open market
half the profit buys DECUPLE at whatever the book quotes, then burns it. the other half is retained.
06 · supply
burned is gone
what is burned leaves circulation for good. how much is whatever step 04 turned out to be.
06 feeds 01 · no paid calls, no loop
profit · the definition
profit = revenue from paid calls
+ subscription revenue
- upstream inference cost
- active development cost
50% of profit → open-market buy, then burn
50% of profit → retained. runs the desks
# no subscriptions are sold yet.
# revenue is the top line, profit is what is left.
# the split is on what is left.
path one · policy
50% of profit, bought on the open market
the other 50% is retained to run the desks and fund the next build. this path moves only when there is profit to split.
path two · mechanical
20 bps on every wDECUPLEx transfer
0.20% of the amount moved. half of that fee burns permissionlessly. it never touches profit and it waits on nobody's decision.
scenario math · dials, not facts
none of this has happened. what follows is arithmetic on numbers we picked, shown in public so you can check the picking. it is a model, not a forecast and not a promise, and it says nothing about what a token is worth. move any input and the output moves with it.
not a forecast · 0 captured · 0 burned
the reference point · public figures
# reported by someone else. not ours, not a partner. reference_rev = ~$140M / yr take_rate = ~5% implied_spend = ~$2.5B - $2.8B / yr decuple_today = 0% of it # implied_spend = reference_rev / take_rate
the dials · we picked these
# assumptions. turn one and the answer moves. capture_share = 5% of that spend savings_ratio = 5x - 10x # demo band, not a promise gateway_margin = 50% - 60% # of what users pay dev_spend = $2M - $4M / yr buyback_share = 50% of profit # policy, not a projection # margin is what users pay, less upstream, plus our cut
Chain
| Step | Scenario value |
|---|---|
| reported revenue, the reference | ~$140M / yr |
| ÷ take rate ~5% | ~$2.5B - $2.8B / yr spend |
| × capture 5% | ~$125M - $140M / yr |
| ÷ savings 5x - 10x | users pay ~$12M - $28M |
| × margin 50% - 60% | ~$6M - $17M / yr |
| less dev $2M - $4M | ~$2M - $15M / yr profit |
| × policy split 50% | ~$1M - $7.5M / yr |
| scenario output, not a forecast. move a dial above and every row under it moves. | |
Sensitivity
| Dial | Which way it pushes |
|---|---|
| savings ratio | widest dial. it sets every row below it |
| context mix | short calls barely save. the band needs long bodies |
| realized pricing | our cut of the saving is policy, not physics |
| dev spend | comes out of profit before the split |
| any one of these can take the last row to zero. | |
the book
the 3x meter only ever sees the slice holographic memory forwards, never the whole body you bound. ids we don't serve resolve to the closest one we do.
421 ids on /v1/models
| model | via |
|---|
353 more · type to thin the list
benchmarks, dev grade
full run: benches.decuple.fun
control run · 1M to 5M
refused
identical bodies · DECUPLE
6/6 HIT
900k
5/6 → 6/6
Test 01
27 / 27 HIT
| Context | Direct | DECUPLE | Used |
|---|---|---|---|
| 60k | 9/9 | 9/9 | 13,102 |
| 200k | 9/9 | 9/9 | 25,088 |
| 500k | 8/9 · 1 UNRUNNABLE | 9/9 | 29,761 |
Test 02
~20k FLAT
| Context | Direct | DECUPLE | Used |
|---|---|---|---|
| 900k | 5/6 · 1 UNRUNNABLE | 6/6 | 19,884 |
| 1M | 0/6 UNRUNNABLE | 6/6 | 19,902 |
| 2M | 0/6 UNRUNNABLE | 6/6 | 20,013 |
| 5M | 0/6 UNRUNNABLE | 6/6 | 20,391 |
due diligence
illustrative quotes · not a live feed
№ 001
https://v1.decuple.fun/v1