# Compute Check by Ashiba Research

## For agents

MCP server (Streamable HTTP, no key): `https://check.ashibaresearch.com/mcp`
- Claude Code: `claude mcp add --transport http compute-check https://check.ashibaresearch.com/mcp`
- Tools: `get_compute_plan`, `get_live_prices`

JSON API. No key.

- `GET https://check.ashibaresearch.com/api/plan?scenario=credits&credits=175000&bill=35000&busy=0.5&shape=agents`
- `POST /api/plan` with the same fields as JSON.
- `GET /api/prices`: live GPU and model API prices.
- Spec: https://check.ashibaresearch.com/openapi.json

Fields:
- `scenario`: credits | bill | renewal | buy | api | lender | idle
- `credits`, `bill`, `api_spend`, `cash`, `burn`, `revenue`: $
- `busy`: 0.01 to 1 (1 = always busy), a percent such as `50%`, or `unknown`
- `shape`: steady | workday | spiky | agents | training | batch
- `expiry_months`, `growth_pct`, `gpu_share_pct`, `gpus_avg`, `gpus_peak`, `mostly_api`, `open_models` (maybe | no)
- Renewal: `contract_gpus`, `contract_end_months`, `contract_rate`
- Idle: `gpu_type` (H100, H200, A100, L40S, RTX 5090, RTX 4090, RTX 3090), `gpus`, `ownership` (owned | contract | cloud), `host_fee_pct`, `power_kwh`
- `team_can_run_servers`, `cloud`, `readiness` (explore | evidence | ready)
- Prices: `rHyper`, `rOd`, `rRes`, `srvPrice`, `colo`, `opsCost`, `life`, `resale`, `minRun`, `loanRate`, `loanTerm`, `loanDown`. Without `rRes`, reserved is 79% of on-demand.

Each answer has `human_url`: the same plan on the page. `rows` is the plan table. Out-of-range values return 400 with `errors`. Quiz buttons carry `data-q` and `data-value`.

Measure GPU use: `curl -O https://check.ashibaresearch.com/gpu_use.py && python3 gpu_use.py --hours 168`, then drop `gpu_use.csv` on the page.

## Your data

- Files stay in your browser. The plan is worked out on your computer.
- Test it. Load the page, turn off wifi. It still works.
- We read only dates, GPU type, hours, cost, and credits. Not account IDs, names, or tags.
- Or skip the bill. Run a query below in your own account and upload the summary.
- Anonymous clicks. We count screens and taps with a random ID. No names, emails, numbers you type, or files. Do Not Track turns it off.
- Leads. Only if you ask: name, email, company, and the plan numbers.

AWS query (Athena)
Replace cur_table. Save as CSV.
SELECT date_trunc('hour', line_item_usage_start_date) AS usage_hour,
product_instance_type AS instance_type,
SUM(line_item_usage_amount) AS hours,
SUM(CASE WHEN line_item_line_item_type <> 'Credit' THEN line_item_unblended_cost ELSE 0 END) AS cost,
SUM(CASE WHEN line_item_line_item_type = 'Credit' THEN -line_item_unblended_cost ELSE 0 END) AS credits
FROM cur_table
WHERE line_item_usage_start_date >= date_add('month', -3, date_trunc('month', current_date))
GROUP BY 1, 2
ORDER BY 1;
Copy

Google Cloud query (BigQuery)
Replace the table name. Save as CSV.
SELECT TIMESTAMP_TRUNC(usage_start_time, HOUR) AS usage_hour,
sku.description AS instance_type,
SUM(usage.amount_in_pricing_units) AS hours,
SUM(cost) AS cost,
-SUM(IFNULL((SELECT SUM(c.amount) FROM UNNEST(credits) c), 0)) AS credits
FROM `project.dataset.gcp_billing_export_v1_XXXXXX`
WHERE usage_start_time >= TIMESTAMP(DATE_SUB(DATE_TRUNC(CURRENT_DATE(), MONTH), INTERVAL 3 MONTH))
GROUP BY 1, 2
ORDER BY 1;
Copy

## Measure GPU use
Run on each GPU machine. It samples nvidia-smi and writes gpu_use.csv. No network calls, no hostnames.
curl -O https://check.ashibaresearch.com/gpu_use.py
python3 gpu_use.py --test
nohup python3 gpu_use.py --hours 168 &
Copy
After a week, drop every gpu_use.csv here. Read the script first.