# GPU9D — independent AI connection quickstart
Status: **research-only**. Real verified GPU offers: **0**. GPU bookings, VM sessions and purchases: **unavailable**.

## Quickest lightweight query (no account)
`GET https://gpu9d.com/api/v1/fast-index?q=G1V24576RNLT3600Xcuda`

The 21-character compact task code represents: GPU count = 1; minimum per-GPU VRAM = 24576 MiB; GPU country = NL; required execution = 3600 seconds; runtime = CUDA. This is a **task filter**, not a physical device ID, hardware validation or available quote. Results currently have zero supplier records.

## MCP (2026-07-28 stateless Streamable HTTP, research only)
Endpoint: `https://gpu9d.com/api/mcp`
Tools: `gpu9d_status` and `gpu9d_index_lookup`.
Each JSON-RPC request supplies headers `MCP-Protocol-Version: 2026-07-28`, `Mcp-Method`, and `Mcp-Name` for tool calls. Every request includes `params._meta` with the protocol version and client capabilities.
```json
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"gpu9d_index_lookup","arguments":{"q":"G1V24576RNLT3600Xcuda"},"_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientCapabilities":{}}}}
```
HTTP header for that example: `Mcp-Method: tools/call`, `Mcp-Name: gpu9d_index_lookup`. Discovery: `server/discover` or `tools/list`. Do not send workload content, credentials or personal information.

## A2A (v1.0, narrow single-turn research)
Agent Card: `https://gpu9d.com/.well-known/agent-card.json`
Endpoint: `https://gpu9d.com/api/a2a`
Use JSON-RPC `SendMessage` with a single `ROLE_USER` text part containing only `status` or a compact task code. No natural-language job requests or booking.

## Machine discovery
- `https://gpu9d.com/ai-index.json` — truthfully describes capabilities and missing services
- `https://gpu9d.com/capabilities.json` — machine feature flags
- `https://gpu9d.com/openapi.json` — HTTP API contract
- `https://gpu9d.com/.well-known/ard.json` — machine discovery metadata

GPU9D currently provides **free research tools only**, not free compute. Professional datacenter resources must be independently verified and explicitly authorized before any real task or commercial connection. No customer datasets, prompts, model weights, credentials or workload commands should be submitted. Anonymous commercial counterparties still require legitimate identity and compliance checks when applicable.
