Compute

#GET /api/compute

Every compute_reward transaction: one per settled hour, from compute.mlx to each wallet whose Macs worked that hour, with each payout's amount and unlock time. It also checks the books: the pool's allocation less its balance must equal the sum of the rewards listed. Newest first.

curl https://mlx.app/api/compute
curl "https://mlx.app/api/compute?format=csv" -o mlx-compute-rewards.csv

const d = await (await fetch("https://mlx.app/api/compute")).json();
console.log(d.reconciled ? "books balance" : "not reconciled", Number(BigInt(d.paid)) / 1e6, "MLX mined");
for (const r of d.rewards.slice(0, 3)) console.log(new Date(r.hour).toISOString(), r.payees, "wallets", Number(BigInt(r.total)) / 1e6);

import requests

d = requests.get("https://mlx.app/api/compute").json()
print("reconciled:", d["reconciled"], int(d["paid"]) / 10**6, "MLX mined")
for r in d["rewards"][:3]:
    print(r["hour"], r["payees"], int(r["total"]) / 10**6)

{
  "account": "compute.mlx",
  "unit": "MLX base units (1 MLX = 1,000,000)",
  "allocated": "3500000000000000",
  "balance": "3499999503946287",
  "paid": "496053713",
  "paidByTransactions": "496053713",
  "transactions": 223,
  "reconciled": true,
  "complete": true,
  "rewards": [
    {
      "hash": "153b2a7cb019ce2a2f64730635c60c564dc096a7844d10979abe062c91198f42",
      "height": 399,
      "hour": 1793113200000,
      "ts": 1793116923936,
      "total": "4298294",
      "payees": 15,
      "payouts": [
        { "to": "luca.mlx", "amount": "31306", "until": 1795651200000 }
      ]
    }
  ]
}

rewards is trimmed to one hour and its payouts to one wallet.

The CSV has the columns tx_hash,block,hour_utc,payee,amount_mlx,unlocks_utc, with times in ISO 8601 and amounts in whole MLX with 6 decimals:

tx_hash,block,hour_utc,payee,amount_mlx,unlocks_utc
153b2a7cb019ce2a2f64730635c60c564dc096a7844d10979abe062c91198f42,399,2026-10-27T15:00:00.000Z,luca.mlx,0.031306,2026-11-26T00:00:00.000Z

Each payout's weight and score are in the transaction's own body (body.payouts[]: to, amount, weight, score): read it with /api/tx/{hash}. How weights and scores are worked out is on Rewards.

#GET /api/compute/work

What the network's work is, in numbers anyone can check: how many tokens Macs produced over the last week and what share of them served a paying request, how often Macs give identical answers to the same job (same chip and across chips), and the baseline of the chip speed check.

No parameters.

{
  "at": 1793116790670,
  "paid": {
    "tokens7d": 55914,
    "paidTokens7d": 0,
    "share": 0,
    "note": "Paid jobs aren't live yet: every token so far is verification work.",
    "days": [
      { "day": "2026-10-27", "tokens": 55914, "paid": 0 },
      { "day": "2026-10-26", "tokens": 0, "paid": 0 }
    ]
  },
  "agreement": {
    "sameChip": { "same": 73, "diff": 0 },
    "crossChip": { "same": 329, "diff": 66 },
    "sameGroup": { "same": 129, "diff": 0 },
    "crossGroup": { "same": 273, "diff": 66 },
    "groups": "M1 | M2 | M3–M4 | M5 and newer. …",
    "pairs": [
      { "a": "M5", "b": "M6", "same": 56, "diff": 0, "agree": 1 },
      { "a": "M6", "b": "M6", "same": 55, "diff": 0, "agree": 1 }
    ]
  },
  "speedCheck": { "baseline": null, "baseMacs": 0, "floor": 0.6, "minSamples": 40, "minBaseMacs": 3 },
  "jobs": {
    "stream": "v2",
    "note": "Each job's prompt is drawn from a seed only the sequencer holds; Macs handed the same job get the same prompt."
  }
}

paid.days and agreement.pairs are trimmed to two entries here.

#paid

These counters started when this endpoint did, so days before it read 0. For the network's tokens today as the Macs report them, see tokensToday on World.

#agreement

When enough Macs have answered the same job, every pair of them is compared once: did they produce identical tokens? Greedy decoding agrees within a GPU generation group but not always across groups, and this measures it.

How answers are compared and a job's verdict is reached is on Jobs.

#speedCheck

A Mac that says it has a Pro, Max, Ultra or newest chip but runs full jobs much slower than base Macs is paid as a base Mac. The check only ever lowers a Mac's class, never raises it.

jobs.stream names the job stream being handed out; jobs.note says how its prompts are made.

#Errors

Neither endpoint takes input that can be wrong: an unknown format returns JSON.