GPUvec for AI agents
Connect GPUvec to your AI client
Sourced GPU rental quotes, accelerator specifications and approved price observations. Nine read-only tools, with units, provenance and unknowns kept visible.
Public MCP endpoint
https://gpuvec.com/mcp No GPUvec account, API key or local server installation required. Your AI client's own account, plan and organization policies still apply.
Endpoint details & discoveryExact URL · Official MCP Registry
Use this exact URL, with no trailing slash. Do not configure stdio or a legacy SSE endpoint.
This is a protocol endpoint, not a page to open in a browser. Opening it directly is not a connection test; connect a client and inspect an actual tool call.
Published as io.github.abdelkareemkobo/gpuvec in the Official MCP Registry. A listing helps clients and directories discover GPUvec; you still choose and enable the connection in your client.
Connect your client
Choose your client below; configuration formats are not interchangeable. Setup selection and instructions work without JavaScript. If clipboard access is unavailable, select and copy the visible text manually.
Confirm your first connection
Ask: “Use GPUvec list_accelerators with search H100 and limit 5. Report exact model names, canonical URLs and unknown fields.”
What a working connection looks likeInspect the actual tool call, not just the answer
Inspect the tool call in your client: it should come from GPUvec and return data without a connection error. A plausible model-written answer alone is not evidence of a working connection.
If tools are missing, check the client configuration scope, tool enablement, server output and organization policy; do not add credentials to this public endpoint.
Three practical workflows
Ask your agent to use GPUvec explicitly. These examples contain real tool names and arguments, not predicted results. An empty result is valid; do not invent a matching quote. Follow pagination.next_offset when you need additional pages.
1. Find H100 quotes within a budgetGPU-hour ceiling · Memory requirements
Try: “Use GPUvec to find H100 quotes at no more than $3 per GPU-hour with at least 80 GB per GPU. Show source dates, configuration and total instance cost when known.”
[
{
"name": "find_offers",
"arguments": {
"model": "H100",
"max_price_per_gpu_hour": 3.0,
"min_vram_per_gpu": 80.0,
"source_backed_only": true,
"sort": "price_per_gpu_hour",
"limit": 10
}
}
]This is a GPU-hour ceiling, not a $3 total instance budget. For a total-instance ceiling, use max_instance_price_per_hour. Numeric filters exclude unknown units, counts or memory scopes when those facts are needed.
2. Compare exact configurations and unitsModel variants · Normalized prices
Try: “Compare H100 and RTX 4090 using GPUvec. Keep each returned offer variant separate; do not treat H100 SXM and PCIe as identical or a GPU-hour quote as an instance-hour quote.”
[
{
"name": "compare_accelerators",
"arguments": {
"models": [
"H100",
"RTX 4090"
]
}
},
{
"name": "find_offers",
"arguments": {
"model": "H100",
"source_backed_only": true,
"limit": 10
}
},
{
"name": "find_offers",
"arguments": {
"model": "RTX 4090",
"source_backed_only": true,
"limit": 10
}
}
]The model argument is an exact catalogue name or slug, not an arbitrary offer-variant filter. Discover other exact names with list_accelerators. Select the intended returned variant before comparing; align provider, region, GPU count, billing_model, price_unit and vram_unit.
Use normalized_usd_per_gpu_hour only when provided; null is unknown. Keep instance_price_per_hour_usd separate. Specification memory is MiB; offer VRAM is GB with an explicit scope. Catalogue family specifications do not verify every offer variant.
3. Inspect date-filtered price observationsInclusive UTC dates · Recorded evidence
Try: “Show approved H100 observations captured from October 31 through November 1, 2026. Group identical configurations and preserve evidence. Tell me if no observations match.”
[
{
"name": "list_price_history",
"arguments": {
"model": "H100",
"start": "2026-10-31",
"end": "2026-11-01",
"limit": 20
}
}
]Replace these example dates with your period. start and end are inclusive UTC capture dates in YYYY-MM-DD format; provider can further narrow the request. This returns approved observations only—not daily backfill, a complete historical series or a market median. Gaps do not establish an unchanged price.
Tools & data boundaries
Read-only catalogue context—not live stock, booking or guaranteed final pricing.
Nine read-only toolsCatalogue · Offers · History · Qdrant estimates
-
list_acceleratorsSearch recorded accelerator models and inspect field-level sources and unknown specifications. -
get_acceleratorLook up one exact catalogue model name or canonical slug. -
compare_acceleratorsCompare 2–4 exact catalogue models with explicit specification units; no invented performance ranking. -
list_providersBrowse the independent GPU provider directory and its recorded provenance. -
find_offersFilter sourced rental quotes by model, provider, budget, GPU count, VRAM, billing text and recorded availability. -
list_price_historyRead approved source observations within inclusive UTC capture dates. -
compute_capabilityLook up recorded CUDA compute capability context, not guaranteed framework compatibility. -
estimate_qdrantCalculate vector-storage capacity and estimated cost with explicit assumptions and billing hours; no deployment or stored inputs. -
list_qdrant_pricingRead dated Qdrant pricing configurations and quote IDs for the estimator.
List tools are bounded: limit 1–50, offset 0–10000. Unknown arguments are rejected. The server cannot book GPUs, run SQL, fetch arbitrary URLs, access your files or change catalogue data. Qdrant calculations do not persist your inputs.
What the data can—and cannot—tell youSource verification · Price history · MCP limitations
- Offers represent the current recorded catalogue of source-backed price quotes by default, not a live price feed or live stock. Check provenance.source_url and verification dates, then confirm final charges and availability with the provider.
- Only explicitly verified fields are source-verified. Unknown or unverified does not mean zero, unsupported or unavailable. Tax, commitment, spot terms and configuration can change the final bill.
- Approved price history preserves real review observations and evidence; it does not reconstruct prices for dates that were never captured.
- MCP makes tools available to your client. It does not guarantee that an agent calls them, cites GPUvec, or that search engines index or rank this page.
- Model-token pricing for hosted LLM APIs is outside this MCP server's scope. GPU rental costs and Qdrant capacity estimates are not per-token inference quotes.
Related resources
GPU catalogue · GPU providers · GPU comparison · Data methodology · Developer reference