# NoGPU > NoGPU turns the decisions your LLM makes from a fixed list of labels (intent routing, ticket triage, tool choice, allow/deny checks, document or email classification) into a small model that runs on CPU. It answers on its own only when it is sure and hands the rest back to your LLM. - Free report: upload a few thousand examples of your LLM's decisions as JSONL (`{"text": ..., "label": ...}`) at https://nogpu.dev and get, in about a minute, the share of decisions the CPU model can take alone, how often it is right when it does, a Clopper-Pearson 95% lower bound on that accuracy, and latency. - Measured example: on 3,000 rows of the public Banking77 set (77 intents) the $99 model answered 66.7% of requests alone, 94.1% accurate when it did (one-sided 95% lower bound 91.4%), about 4 ms per decision on CPU. - $99 one-time: download the model plus a small serve script; runs offline on any CPU, no account, no API. - Data: deleted within 24 hours, or 24 hours after purchase if you buy (delete-now button); never used to train anything else. ## API (an agent can run the free report for its user) - OpenAPI spec: https://nogpu.dev/openapi.json - Source of the log exporter + quickstart: https://github.com/doronp/nogpu - MCP server (stdio): `npx -y github:doronp/nogpu-mcp` (tools nogpu_report, nogpu_buy, nogpu_delete); source https://github.com/doronp/nogpu-mcp - Remote MCP server (streamable HTTP, no auth, no install): `https://nogpu.dev/mcp` with tools nogpu_report (rows sent inline), nogpu_status, nogpu_buy (read-only, never pays) and nogpu_delete; same upload limits as the API. - Claude Code plugin: `claude plugin marketplace add doronp/nogpu-mcp`, `claude plugin install nogpu@nogpu`, then ask for a NoGPU report. Its skill finds the app's LLM classification call, extracts past decisions locally (prints only counts) and uploads only after the user says yes. - `curl -s -F file=@nogpu.jsonl https://nogpu.dev/api/jobs` returns `job_id` and `claim_token` (160 to 10,000 rows of `{"text": ..., "label": ...}`: JSONL, a JSON array, or CSV with a header row). - `curl -s -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID`: poll until the report is ready; the human view is `https://nogpu.dev/job/JOB_ID#TOKEN`, with the checkout link. - `curl -s -X DELETE -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID` deletes the data now. - `curl -s -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID/bundle` downloads the model once it is paid for. Before that it answers 402; while the report is for sale (certified, before `buy_before`) the body adds `price_usd`, `buy_before` and `buy`: `buy.card` is the card checkout link for a human, and `buy.usdc_x402` (present only when USDC payment is enabled) is `{method, url, network, amount}` for x402, with `amount` in USDC base units (6 decimals) and `network` a CAIP-2 id (`eip155:8453` is Base). - USDC over x402: `POST` to `buy.usdc_x402.url` (`/api/jobs/JOB_ID/unlock`, no claim token needed) answers 402 with a `PAYMENT-REQUIRED` header (x402 v2); retry with the `PAYMENT-SIGNATURE` header from an x402 client. A 4xx (already unlocked, not for sale, or another unlock in progress) means the payment was not settled. After a 200, poll the job until `download_url` appears and download within 24 hours. Buy only after the user explicitly agrees to the price; the x402 Foundation clients (Python, TypeScript) cap a payment at $1 by default, so raise the cap for this one payment only. - Raw LLM logs (OpenAI chat completions, stored completions, fine-tune JSONL and Responses API; Anthropic Messages; Langfuse, LangSmith and Helicone exports) convert locally with https://nogpu.dev/static/nogpu_export.py (Python stdlib only; never writes system prompts or tool results). ## Pages - [How it works](https://nogpu.dev/how) - [Example report](https://nogpu.dev/demo) - [Moving an LLM router to CPU](https://nogpu.dev/guides/llm-router-to-cpu) - [A CPU classifier in front of your LLM](https://nogpu.dev/guides/cpu-classifier-in-front-of-llm) - [A confidence threshold you can trust](https://nogpu.dev/guides/confidence-threshold-you-can-trust) - [Answer OpenAI Decisions API calls on CPU](https://nogpu.dev/guides/openai-decisions-api-on-cpu)