Model Context Protocol server

Stop paying frontier prices for cheap work.

Classification, summarization, extraction, PII handling, and short chat turns do not need a frontier model. Point your agent here and they run on ZeroGPU's small and nano models instead — and every result tells you exactly what it cost, and what it would have cost.

Connect a client Browse the tools

Endpoints

Everything but the landing page, the server card, and the health probe requires a ZeroGPU API key.

/mcp
Streamable HTTP — the recommended transport.
/sse
Server-Sent Events, for clients that need it.
/health
Liveness and environment, no credential required.
/.well-known/mcp.json
Server card — name, version, and tool list.

Twelve tools

Each one is task-centric: it names the job rather than the model, so a calling agent picks by what it is trying to do.

zerogpu_classify_iab

Classify a passage against the IAB Content Taxonomy. Pass enriched for the richer label set.

zlm-v1-iab-classify-edge
zerogpu_classify_domain

Map a bare hostname to categories, topics, keywords, and user intent — no page fetch.

zlm-v1-iab-domain-classifier
zerogpu_summarize

Condense any passage. Returns a bare summary with no preamble, safe to use verbatim.

llama-3.1-8b-instruct-fast
zerogpu_classify_zero_shot

Score text against a flat list of labels you supply. No training data, no fixed taxonomy.

deberta-v3-small
zerogpu_classify_structured

Classify one passage along several named axes at once, each with its own label set.

gliner2-base-v1
zerogpu_extract_entities

Named-entity recognition against labels you define — including ones no standard taxonomy has.

gliner2-base-v1
zerogpu_extract_json

Pull named fields out of free text into a JSON object shaped by your own schema.

gliner2-base-v1
zerogpu_redact_pii

Mask personal information in place before you log, share, or forward the text.

gliner-multi-pii-v1
zerogpu_extract_pii

Report what personal data a passage holds, grouped by category, leaving it unchanged.

gliner-multi-pii-v1
zerogpu_generate_followups

Propose the questions a reader would ask next about a passage.

zlm-v1-followup-questions-edge
zerogpu_chat

A general chat turn, with an optional reasoning trace and a model override for heavier work.

LFM2.5-1.2B-Instruct
zerogpu_health

Confirm the backend is reachable and read per-component status. Runs no model.

no model

Connect a client

Pass your ZeroGPU API key as an x-api-key header or an ?apiKey= query parameter. A project ID is optional — the backend derives the project from the key.

01 / CLAUDE CODE

Add the remote server

claude mcp add --transport http zerogpu \
  https://mcp.zerogpu.ai/mcp \
  --header "x-api-key: YOUR_ZEROGPU_API_KEY"
02 / ANY MCP CLIENT

Or use the query parameter

https://mcp.zerogpu.ai/mcp?apiKey=YOUR_ZEROGPU_API_KEY

Useful for clients that cannot set custom headers. Prefer the header where you can — query strings end up in logs.

03 / VERIFY

Check it answered

curl -s https://mcp.zerogpu.ai/health

Then call zerogpu_health from your client to confirm the key itself works.

On every result

The bill, itemized

Every tool returns a savings block next to its answer: the tokens consumed, what they cost on ZeroGPU, what the same tokens would have cost on a frontier model, and the version of the price table used to work it out. 

"savings": {
  "input_tokens": 812,
  "output_tokens": 96,
  "zerogpu_cost_usd": 0.00002104,
  "baseline_cost_usd": 0.003876,
  "savings_usd": 0.00385496,
  "price_table_version": "2026-08-21"
}