Use case · Agents

Agents burst. Idle is free.

Agent traffic arrives in bursts: a plan, a run of tool calls, then silence. Serverless endpoints add workers for the burst and drop to zero in the silence, billed per second either way.

Get started Talk to sales

H200 · €0.00111 /GPU·sidle at €0.00one key per agent

traces / support-agent

streaming

run_8f21c · 8 steps€0.0050

plan
1100ms
search_docs
rerank
300ms
draft_reply
1300ms
run_code
check
700ms
update_ticket
summarise
1100ms

GPU tooltotal 6.9 s · 4.5 GPU-s

Teams running agents on Nexpod

Placeholder

Customer
CUSTOMER
customer
Customer Co
customer.ai
CUSTOMER

Customer stories

Agents doing real work.

Coding agentH200

Customer

PlaceholderServerless

Placeholder quote. A customer explains, in two plain sentences, what changed once their coding agent workload moved to Nexpod.

Customer nameRole, Company

Read case study

Support agentH100

Customer

PlaceholderServerless

Placeholder quote. A customer explains, in two plain sentences, what changed once their support agent workload moved to Nexpod.

Customer nameRole, Company

Read case study

Research agentB200

Customer

PlaceholderCloud Servers

Placeholder quote. A customer explains, in two plain sentences, what changed once their research agent workload moved to Nexpod.

Customer nameRole, Company

Read case study

Browser agentH200

Customer

PlaceholderServerless

Placeholder quote. A customer explains, in two plain sentences, what changed once their browser agent workload moved to Nexpod.

Customer nameRole, Company

Read case study

Nexpod Hub

Start from an agent-ready template.

Hub templates bundle a model server, scaling defaults and a health check. Deploy one in a click, then swap in your own weights or prompts.

Browse Nexpod Hub

hub/qwen3-agent

vLLM · tool calling · chat route

Popular

H200 · from €0.00111/GPU·sDeploy

hub/llama-tools-70b

TGI · function calling

B200

hub/embed-small

text embeddings · batch 256

H100

01Bursts

Zero workers to forty, and back.

Set a minimum and a maximum per endpoint. Workers follow queue depth, so a spike of tool calls finds capacity you did not have to book.

Queue-based scaling: Scale on queue depth or on requests per worker, per endpoint.
Idle timeout: Choose how long a worker stays warm after its last request.

workers · endpoint agent-llm · one weekday · min 0 · max 40

plan · 12

burst · 40 workers

02Sessions

Long tasks keep their GPU.

Some agents work for minutes, some for hours. Pin a session to a dedicated worker or a Cloud Server so its state and caches stay where they are.

Dedicated workers: Reserve workers for sessions that must not be moved mid-task.
Scoped API keys: Give each agent its own key with a daily limit in euros.

sessions · team agents

3 active

AgentTimeState

code-review1h 18mrunning

research42m 10srunning

support3m 04srunning

support58sidle

data-clean2h 06mdone

pinned sessions keep their worker

03Tools

Models and tools in one region.

Keep the model, the embeddings, the code sandbox and the vector store in one region, so each step skips the trip across the internet.

Private network: Endpoints, servers and volumes in a region share one private network.
Isolated sandboxes: Run untrusted tool code in separate containers, away from your model.

EU-FRA-1 · private network

agent-apiyour orchestrator

llm-euendpoint · 8× H200

0.6 ms

embed-euendpoint · H100

0.5 ms

sandbox-poolisolated containers

0.8 ms

vol/vectorsnetwork volume

0.4 ms

round trip per step, same regionegress inside region €0.00

Developer tools

Call it from any agent framework.

Endpoints speak plain HTTPS and offer a chat-completions route for chat models, so most frameworks only need a base URL and a key to use them.

Get started Read the docs

1import nexpod
2 
3client = nexpod.Client()
4llm = client.endpoints.get("agent-llm")
5 
6step = llm.run({
7    "messages": history,
8    "tools": tools,
9    "session": "ses_41a9",  # keeps this run on one worker
10})
11 
12history.append(step.output)