NoGPU
● Activeio.github.doronp/nogpu
Free report: how many of your LLM's classification calls a small CPU model can answer alone.
Connect this server
One click to install, or copy the config below. Suits: Cursor · HTTP clients.
Or add it manually — claude_desktop_config.json (Claude Desktop) or .cursor/mcp.json (Cursor):
{
"mcpServers": {
"nogpu": {
"url": "https://nogpu.dev/mcp"
}
}
}claude mcp add --transport http nogpu https://nogpu.dev/mcpcode --add-mcp '{"name":"nogpu","type":"http","url":"https://nogpu.dev/mcp"}'Syntax from each client's docs (Claude Code: claude mcp add; Codex: codex mcp add; VS Code: code --add-mcp). Placeholders in angle brackets are values you fill in. Codex takes hosted servers in config.toml ([mcp_servers.<name>] url = …) rather than on the command line.
A starting point — check the repository for exact arguments and setup. Restart your client after editing the config.
- Category
- AI & Agents
- Version
- 0.2.3
- Registry
- active
- Published
- 11/18/2026
Is it safe? Is it maintained?
Serverdex check: Nov 22, 2026Looks healthy: NoGPU is maintained, open source, hosted endpoint answering without a key.
- Maintained
- Active
- Last commit 4 days ago (Nov 18, 2026), 0 open issues.
- Official registry
- Active
- Published as io.github.doronp/nogpu in the official MCP registry, version 0.2.3.
- Source
- Open source, MIT
- Public repository under the MIT licence, 0 stars, 0 forks.
- Runs
- Local package or hosted
- Ships as a local package (mcpb) that runs on your machine, and as a hosted endpoint at nogpu.dev that receives what your client sends it.
- Credentials
- None needed to connect
- The hosted endpoint completed the Serverdex initialize + tools/list handshake with no key (checked Nov 22, 2026); individual tools may still ask for one.
- Owner
- Unclaimed
- Nobody has claimed this page yet; the data above comes from the registry, GitHub and the live check.
Facts only: last commit and licence from GitHub, the entry's status from the official registry, credentials from the live check. Not a code audit — run the security scan on the endpoint yourself and see the status history.
Live check
checked Nov 22, 2026, 00:07 UTCStatus history & uptime →The endpoint answered the MCP handshake as "NoGPU" and listed 4 tools. Answering since Nov 18, 2026.
Context cost ≈ 932 tokens of tool definitions a client loads per conversation (estimate: characters ÷ 4) across 4 tools.
Tools (4)
nogpu_report4 paramsFree report: send an LLM's past decisions, 160-10,000 rows with at least 2 labels, either as `rows` ([{"text": ..., "label": ...}]) or as `jsonl` (the same rows as JSONL text; a JSON array or a CSV with a header row also works), at most 20 MB. NoGPU trains a small CPU model on them and measures what share of the decisions it answers alone, how often those answers match the LLM (with a 95% lower bound) on held-out rows, and its latency. Waits up to 40 seconds; if the report is not ready by then, call nogpu_status with the returned job_id and claim_token. claim_token and report_url (the token is after its #) open the job: treat both as secrets. Free reports through this endpoint come from one shared daily pool. Data is deleted within 24h. `source` says where the rows came from: agent (default), the app's stored history, rows logged by a collect snippet, or a given file. No data at hand? Set `sample` instead of rows/jsonl to run a real report on a public set (3,000 rows; clinc150 and massive-en are assistant intent routing, civil-comments-toxic is moderation): see https://nogpu.dev/benchmarks.
Show parametersHide parameters
rowsjsonlsourcestringsample
nogpu_status2 paramsA job's status, and its report once ready (same fields as nogpu_report). Also shows download_url after payment: GET it with header X-Claim-Token. claim_token and report_url are secrets.
Show parametersHide parameters
job_idstringrequiredclaim_tokenstringrequired
nogpu_buy2 paramsRead-only, never pays: for a certified (fit) report, returns the $99 offer to keep the model, the question to ask the user word for word and a card checkout link for the user. If the job is not buyable, returns the reason.
Show parametersHide parameters
job_idstringrequiredclaim_tokenstringrequired
nogpu_delete2 paramsDelete an uploaded job's data now (otherwise deleted within 24h, or 24h after payment if you bought the model).
Show parametersHide parameters
job_idstringrequiredclaim_tokenstringrequired
Checked by Serverdex with a plain initialize + tools/list call to nogpu.dev/mcp, no credentials, about once a day. A failed check is never a removal.
On GitHub
read Nov 18, 2026- Stars
- 0
- Forks
- 0
- Last push
- Nov 18, 2026
- License
- MIT
Pythonclassificationclaude-codeclaude-code-plugincost-optimizationdistillationllmllm-routermcpmcp-server
From the README
nogpu-mcp MCP server that lets your agent measure, on your own data, how many of your LLM's classification or routing decisions a small CPU model can take alone. It calls the free NoGPU report. Tools: • nogpureport(path, source?, callsperday?, usdper1kcalls?): uploads a file of {"text": ..., "label": ...} rows (JSONL, a JSON array, or a CSV with text and label columns; 160 to 10,000 rows, at least 2 labels) and returns, usually in 15-60 seconds: a fit / not-fit verdict, the share the CPU model answers alone, how often it then matches your LLM with a one-sided 95% Clopper-Pearson lower bound (held-out rows), latency, the row counts, what the upload check changed, and a link to the report. Keeping the model is $99 one-time per model (a new label set is a new upload and a new model); the report is free. source (history, logging or file, the default) is sent as the X-NoGPU-Source header so NoGPU can count where reports come from. callsperday and usdper1k_calls (your estimate) are added to the report link so its savings calculator starts from your numbers. • nogpubuy(jobid, claim_token): read-only, never pays. For a fit report it returns the $99 offer, the yes/no question to ask you, a card checkout link for you to open, and the USDC (x402) payment details when the server offers them; for any other job, the reason it can't be bought.…
Read the full README on GitHub →Remotes
- streamable-http
https://nogpu.dev/mcp
Alternatives to NoGPU
All NoGPU alternatives →Best AI & Agents servers →The most-starred AI & Agents servers in the index, with their live-check result.
| Server | Stars | Last push | Live check |
|---|---|---|---|
| Ruflo | 81,994 | Nov 2026 | Local / not checked |
| Claude Flow | 81,985 | Nov 2026 | Local / not checked |
| CodeBurn | 11,880 | Nov 2026 | Local / not checked |
| Kaneo | 9,825 | Nov 2026 | Local / not checked |
| Ha MCP | 7,958 | Nov 2026 | Local / not checked |
NoGPU: questions people ask
- Is NoGPU a remote or a local MCP server?
- According to its entry in the official MCP registry, NoGPU ships as a hosted endpoint (streamable-http) that your client connects to by URL and as a local package (mcpb) that your client starts on your machine. Use the hosted endpoint when your client supports Streamable HTTP; use the package when you want it to run locally.
- Does NoGPU need an API key?
- Its hosted endpoint completed the Serverdex live check — an MCP initialize + tools/list call — with no credentials on Nov 22, 2026. Individual tools may still ask for a key when called.
- Is NoGPU safe to use?
- Looks healthy: NoGPU is maintained, open source, hosted endpoint answering without a key. In detail — maintained: Active; official registry: Active; source: Open source, MIT; runs: Local package or hosted; credentials: None needed to connect. Serverdex check of Nov 22, 2026. This is a reading of the repository, the registry entry and the live check, not a code audit; run the Serverdex security scan on the endpoint before trusting it with real data.
- Is NoGPU still maintained?
- Active: the repository's last commit 4 days ago. Last push Nov 18, 2026, 0 GitHub stars, MIT licence. Read from GitHub on Nov 18, 2026.
- What tools does NoGPU expose?
- 4 tools as listed by the server itself during the live check, including nogpu_report, nogpu_status, nogpu_buy, nogpu_delete. The full list with parameters is in the Live check section above.
- How do I add NoGPU to Claude, Cursor or VS Code?
- Copy the config in "Connect this server" into your client's MCP config file (claude_desktop_config.json, ~/.cursor/mcp.json or .vscode/mcp.json), or use the one-click Add to Cursor / Add to VS Code buttons and the one-line commands for Claude Code and Codex; then restart the client.
More AI & Agents servers
Browse all AI & Agents →A2A Peptides
Dosidal
Agent-to-Agent (A2A) + Model Context Protocol (MCP) hub for peptides.
Banco MCP
douglac
Brazilian Open Finance MCP — 30+ banks (Itaú, Nubank, etc.) to Claude/Cursor. Read-only.
EMET
dreamforgestudiollc
ISO 9001 7.5.3 / ISO 13485 4.2.5 document-control memory for AI agents. Nothing is deleted.
MemVault - Local-First Agent Memory
dreamor
Local-first memory for AI agents: save, recall, dedupe, and share across MCP clients.
GHL Command
drjerryrelth
GoHighLevel MCP for Claude: 325+ tools, 69 modules, including a workflow builder
Caprock
dspv
Read-only memory of your coding agents: decisions, past sessions, spend and CI. Needs Caprock.
Add the badge to your site
Show visitors that NoGPU is listed here — it’s also the link we check for a free listing.
HTML — for your website
<a href="https://serverdex.io/server/io-github-doronp-nogpu?utm_source=badge"><img src="https://serverdex.io/brand/badge.svg" alt="NoGPU — listed on Serverdex" width="123" height="40" loading="lazy"></a>Recommended. A dofollow link from your own domain is the one that counts for search engines, and it's the link we check for a free listing.
Markdown — for your README
[](https://serverdex.io/server/io-github-doronp-nogpu?utm_source=badge)Nice for visibility, but GitHub marks README links nofollow, so it doesn't pass authority. Use the HTML badge on your site as well.
Install buttons — for your README
[](https://serverdex.io/install/cursor/io-github-doronp-nogpu) [](https://serverdex.io/install/vscode/io-github-doronp-nogpu) [](https://serverdex.io/install/claude-code/io-github-doronp-nogpu)One click opens Cursor or VS Code with this server's config prefilled; the Claude Code button lands on the one-line command. The links point at Serverdex, so when your config changes the buttons keep working.
Install buttons — HTML for your docs
<a href="https://serverdex.io/install/cursor/io-github-doronp-nogpu"><img src="https://serverdex.io/brand/install-cursor.svg" alt="Add to Cursor — via Serverdex" width="140" height="28"></a>
<a href="https://serverdex.io/install/vscode/io-github-doronp-nogpu"><img src="https://serverdex.io/brand/install-vscode.svg" alt="Add to VS Code — via Serverdex" width="147" height="28"></a>
<a href="https://serverdex.io/install/claude-code/io-github-doronp-nogpu"><img src="https://serverdex.io/brand/install-claude-code.svg" alt="Open in Claude Code — via Serverdex" width="180" height="28"></a>Same buttons as plain links for a docs page or website.
Prefer a live badge with the score, stars and live check? Score badge (links to a page that explains the number and back to this listing).