docs
Docs
Send past decisions of one recurring LLM call (a route, a category, allow or deny) and get a free report: the share of those decisions a small CPU model answers alone, how often it matches your LLM when it does, with a 95% lower bound, and its speed. Pick whichever way in suits you.
- Web: drop a file on the home page.
- HTTP API: one
curl, below (API). - Any MCP client: add
https://nogpu.dev/mcp, no install, no key (MCP). - Claude Code plugin: finds the LLM call in your app, extracts its past decisions locally and asks before uploading (plugin).
No data at hand? The example report and the benchmarks show real reports on public sets.
{"text": "where is my card?", "label": "card_arrival"}
{"inputs": {"plan": "pro", "subject": "refund"}, "label": "billing"}
- One row per past decision:
text(or aninputsobject of fields) and thelabelyour LLM chose. Optional:id,source,teacher_probs,teacher_label,split. Other fields are dropped. - JSONL, a JSON array, or CSV with a header row. 160 to 10,000 rows, at most 20 MB, at least 2 labels.
- About a quarter of the rows are held out to measure the model, so fewer than about 160 rows cannot clear the accuracy bar; a few hundred rows per label is better.
nogpu_export.py turns OpenAI (chat completions, stored completions, fine-tune JSONL, Responses
API), Anthropic Messages, Langfuse, LangSmith and Helicone logs into rows. It runs locally on the Python
standard library, sends nothing anywhere, and writes only the last user message and the decision, never
system prompts or tool results.
curl -sO https://nogpu.dev/static/nogpu_export.py
python3 nogpu_export.py logs.jsonl > nogpu.jsonl # auto-detects the format
python3 nogpu_export.py logs.json --input-field input.query --label-field output.route
It prints the row and label counts, any label spellings it merged, labels with fewer than 5 rows, and whether the file is ready to upload; no row text.
# upload: returns job_id, claim_token and job_url (open https://nogpu.dev + job_url)
curl -s -F file=@nogpu.jsonl https://nogpu.dev/api/jobs
# status and report, once status is report_ready
curl -s -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID
# delete the data now
curl -s -X DELETE -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID
# the model, after payment (before that: 402, with the $99 offer while the report is for sale)
curl -s -o bundle.tar.gz -H "X-Claim-Token: TOKEN" https://nogpu.dev/api/jobs/JOB_ID/bundle
The claim token is the only key to a job: keep it secret. Spec: /openapi.json. For agents: /llms.txt.
Remote, streamable HTTP, no install, no key: https://nogpu.dev/mcp.
# Claude Code
claude mcp add --transport http nogpu https://nogpu.dev/mcp
// Cursor: .cursor/mcp.json in a project, or ~/.cursor/mcp.json
{"mcpServers": {"nogpu": {"url": "https://nogpu.dev/mcp"}}}
nogpu_report: send rows inline (rows, orjsonltext) and get the report. Or setsampletoclinc150,massive-enorcivil-comments-toxicto see a real report on a public set first.nogpu_status: a job's status and report.nogpu_buy: read-only, never pays: the $99 offer and a checkout link for the user.nogpu_delete: delete the data now.
Free reports through the remote server come from one shared daily pool. Rows sent as tool arguments
also pass through your agent's model provider. A local stdio server reads a file from disk instead:
npx -y github:doronp/nogpu-mcp (source).
claude plugin marketplace add doronp/nogpu-mcp
claude plugin install nogpu@nogpu
Then ask "get me a NoGPU report for this app". The skill finds the recurring classification or routing
call, extracts its past decisions into nogpu.jsonl, shows only counts and label names, and
uploads after you say yes. With no stored history, it asks before adding a small local logging snippet.
- If the model clears the bar, the report page offers it for $99, one time. The download is the model,
its confidence gate, the report and
serve.py, which answers offline on any CPU, also as a local OpenAI Decisions API endpoint (guide). - Your data is deleted within 24 hours of upload, or 24 hours after purchase. Delete it sooner from the
report page, the API or
nogpu_delete.