Access and inference

A successful reservation does not yet provide execution access. Poll GET /api/v1/jobs/{id} and wait for RUNNING, then request GET /api/v1/jobs/{id}/access.

Choose the right access path

Access	Credential and destination
AmpleRun’s OpenAI-compatible relay	Account API key at https://amplerun-api.spiry.org/api/v1/jobs/{id}/openai/v1
A template’s directly exposed model endpoint	The endpoint and per-rental api_key returned in access details
SSH	Your own private SSH key, matching the public key supplied when reserving

Do not swap the per-rental key and account API key. Keep both out of shared notebook output, shell traces, and application logs.

Call the model relay

With AMPLERUN_API_KEY and AMPLERUN_JOB_ID already provided privately to your process, this read-only Python standard-library example lists the rental’s model IDs:

import json
import os
import urllib.request

url = (
    'https://amplerun-api.spiry.org/api/v1/jobs/'
    + os.environ['AMPLERUN_JOB_ID']
    + '/openai/v1/models'
)
request = urllib.request.Request(url, headers={
    'Authorization': 'Bearer ' + os.environ['AMPLERUN_API_KEY']
})
with urllib.request.urlopen(request, timeout=30) as response:
    models = json.load(response)
print([model['id'] for model in models.get('data', [])])

Use an actually returned model ID for /chat/completions, /completions, or /embeddings where the selected template supports that operation. The relay follows the running rental; it is not a general pool that selects a different model or machine for you.

SSH and result retrieval

Provide a one-line OpenSSH public key when creating a rental if you need SSH. Without it, SSH is disabled; model/API-only use does not require a key in the current REST and hosted MCP flows. Check MCP version differences before using an older local client.

Use the returned ssh_command and verify ssh_host_key_fingerprint. Never turn off host-key verification just to bypass an unexpected mismatch.

After STOPPED, access may return mode: retrieval-read-only until expires_at. That access is for retrieval, not execution. Move important results off the machine before stopping; do not assume a retrieval endpoint exists for every template or tier.