MCP SERVER · OPEN SOURCE

Your agents should get discounts

Validated codes, ranked by what happened at a real checkout.

https://discoball.dev/mcp
—
CODES VERIFIED
—
CHECKOUTS RUN
READ THE SOURCE ↗

The catalog in exact figures

01 / merchants
50

Stores live across 7 categories, from apparel to pet supplies.

02 / verified
13

Codes proven at a real checkout, by hand or by an agent report.

03 / freshness
—

Since the last verification ran against a live cart.

04 / latency
5ms

Median server response on find_deals, p95 under 10 milliseconds.

01 / 03

find_deals

Ask by merchant, URL or category. Returns validated codes with their evidence attached, not a scraped list.

SEE THE SCHEMA
inputmerchant · url · category
returnsranked codes + proof
authnone
02 / 03

rank

Sorts on status, age and observed success rate. Commission is not an input, and a CI test enforces it.

SEE THE SCHEMA
signalsstatus · age · hit rate
excludedcommission
enforced byci test
03 / 03

report_result

Every checkout writes back. Successes hold a code's position, failures drop it within minutes.

SEE THE SCHEMA
inputcode · outcome · amount
effectre-ranks catalog
authnone · installation id
WE DON'T RANK BY COMMISSION. WE RANK BY EVIDENCE.  ·  WE DON'T RANK BY COMMISSION. WE RANK BY EVIDENCE.  · 

Verified right now

order set by evidence
CODEMERCHANT RATESAVED
THANKYOU10allbirds —$14.00
KAT10everlane —10%
NEWVUORI-V220vuori —20%
commission is not an input · test/commission-blind.test.ts

How a code earns its rank

Three inputs, all of them observable. Anything an affiliate program pays us stays out of the sort.

01.

Ingest

Codes arrive from merchant sites and public promotions, normalized into one shape.

02.

Verify at a cart

Codes are run against live checkouts, by hand and by agent reports. Until one passes, a code is labeled unproven — never dressed up as verified.

03.

Score

Offer value, verification age and success confidence combine into one number. Nothing about payout enters here.

04.

Serve

Agents get the ranked set with the evidence behind each position, in a few milliseconds.

05.

Learn

Reported outcomes feed straight back into the score, so the catalog corrects itself between crawls.

FAQ

FOR AGENT BUILDERS

Not to read, and not to write. find_deals is open to any MCP client, and report_result takes a self-issued installation id so we can rate-limit and trace bad data.

Merchant sites and public promotions. Every one is labeled by what happened at a real cart: tested codes read verified, untested ones read unproven until a checkout proves them.

No. The sort reads offer value, verification age and observed success rate only. A CI test fails the build if a commission field ever reaches the ranking function, and the test is public.

The agent reports the failure through report_result, the success rate drops, and the code falls out of the top slot within minutes rather than at the next crawl.

Eventually, affiliate revenue on codes that convert, disclosed per deal. It changes what we earn, never what we surface first.

Point your agent at it

One endpoint, no account. Paste the URL into your MCP client and the tools show up.

GITHUB ↗
TRANSPORTstreamable http · stdio
AUTHnone to read
AGENT DOCS/llms.txt