For agents

Compute Grid is an open, MCP-native price feed for the GPU spot market. Agents can find, compare, and delegate to the cheapest available GPU as easily as they list files. Everything below is free and unauthenticated.

Connect via MCP

One command. No keys, no accounts. Streamable HTTP transport, session-less, JSON responses by default.

claude mcp add --transport http computegrid https://[host]/mcp

The same endpoint works in Cursor, custom Anthropic SDK clients, and any MCP-compatible host.

Tools

list_gpus(filters)

Returns matching GPU rows sorted by price ascending. Every filter is optional.

{
  gpu_model?: string,             // substring match
  provider?: string[],            // e.g. ["runpod", "lambda-labs"]
  tier?: string[],                // ["secure","community","verified",...]
  max_price_per_gpu_hour?: number,
  min_vram_gb?: number,
  available_only?: boolean,       // default true
  region?: string,                // substring match
  limit?: number                  // default 100
}

find_cheapest(input)

One call, full context. Returns the cheapest matching offer plus 5 alternatives plus market context (median, total available, cheapest 24h ago).

{
  gpu_model: string,        // required
  gpu_count?: number,       // default 1
  min_vram_gb?: number,
  tier?: string[],
  region?: string
} -> {
  cheapest: GpuRow,
  alternatives: GpuRow[],
  market_context: {
    median_price_per_gpu_hour_usd: number,
    total_available_count: number,
    cheapest_24h_ago: number | null
  }
}

HTTP endpoints

For agents that don't speak MCP, the same data is open over plain HTTP.

• GET /api/snapshot.json ·full current snapshot, { fetched_at, rows: GpuRow[] }, cached 60s
• GET /api/stream ·Server-Sent Events; snapshot on connect, diff on every refresh (~120s)
• GET /api/health ·liveness + last snapshot age
• GET /llms.txt ·agent-discoverable site description (per llmstxt.org)

Schema (GpuRow)

{
  id: string,                          // stable across refreshes
  provider: string,                    // "runpod", "vast", "lambda-labs", ...
  tier: "secure" | "community" | "verified" | "unverified" | "standard",
  gpu_model: string,                   // canonical: "H100 SXM", "A100 80GB"
  vram_gb: number,                     // per single GPU
  gpu_count: number,                   // 1, 2, 4, 8
  price_per_gpu_hour_usd: number,
  price_per_instance_hour_usd: number,
  available: boolean,
  offer_count: number,                 // > 1 for marketplace dedup groups
  regions: string[],
  metadata: {
    raw_provider_id: string,
    source?: "getdeploying",           // present when sourced via aggregator
    source_url?: string,
    ...                                // provider-specific extras
  },
  fetched_at: string                   // ISO 8601
}

Data sources

Three direct integrations refreshed every 120 seconds: RunPod (GraphQL), Vast.ai (REST bundles), Vultr (REST plans).

Plus 30+ additional providers ·Lambda Labs, CoreWeave, AWS, GCP, Azure, OVH, Scaleway, Crusoe, Hyperstack, Cudo Compute, TensorDock, Paperspace, Fluidstack, and more ·sourced via getdeploying.com, which updates daily upstream. Aggregator-sourced rows are tagged metadata.source = "getdeploying".

Tier semantics

• Managed cloud (secure, standard): listed by managed cloud providers with SLAs and support. Higher prices, predictable.
• Marketplace (community, verified, unverified): peer-to-peer listings (Vast.ai, RunPod Community). Lower prices, more variance, no SLA.

Example: agent shopping for compute

agent → find_cheapest({ gpu_model: "H100 SXM", gpu_count: 8 })

→ cheapest:        $1.33/hr/GPU  · Vast.ai Unverified
  alternatives:    5 within +20%
  cloud price:     $2.99/hr/GPU  · RunPod Secure
  market_context:  32 listings, median $2.49, no 24h history yet

Open and verifiable

Every row links back to its source. No markup, no platform fee, no API key. Source code is on GitHub.