ModelDock local AI benchmark

Paste this address into your AI agent (Claude Code, Codex, Cursor, or any agent that can run commands) and ask it to benchmark your local model. You get a share card with your speeds and where they rank.

中文：把这个网址发给你的 AI 助手，让它帮你测本地 AI 的速度。日本語：このアドレスを AI エージェントに渡して、ローカルモデルの速度を測ってもらってください。

For the AI agent

1. Start a session. No MCP setup is needed:

   curl -s -X POST https://speed.modeldock.net/v1/start -H "content-type: application/json" -d "{}"

   Invoke-RestMethod -Method Post -Uri https://speed.modeldock.net/v1/start -ContentType 'application/json' -Body '{}'

   If your client supports MCP, you can instead connect https://speed.modeldock.net/mcp and call start_benchmark. Pass {"port": N} if you know the local model server's port.
2. Follow the returned instructions exactly: download and verify the reference script, read the hardware from the system, and ask the person before running it (it uses the GPU for 2 to 3 minutes and only talks to the model server on this computer).
3. Machine name: if the person gave one, add it to hardware.json as "share": {"machineName": "…"}. If not, leave it out: the card shows the hardware you read. Never invent a name.
4. When the script has submitted, get the result in the person's language:

   GET https://speed.modeldock.net/v1/sessions/<sessionId>/result?locale=en|zh-CN|ja

   It returns the share link (shareUrl) and the share card as a 1200×630 PNG image (pngUrl), both in that language. If your interface can show images, show the person the card; always give them the share link unchanged.

Full agent guide: https://speed.modeldock.net/guide.md