A GPU tool your agent
can actually use.

One endpoint. Temporary private access.

In a client that supports remote MCP, choose Streamable HTTP and use:

https://u8x5gpk0ul.execute-api.eu-north-1.amazonaws.com/v1/mcp

Connect without an authentication header. The create_trial_access tool issues a 24-hour token. Keep it private and pass it as access_token to job tools. Your client's tool approval settings still apply.

{
  "servers": {
    "agentcompute": {
      "type": "http",
      "url": "https://u8x5gpk0ul.execute-api.eu-north-1.amazonaws.com/v1/mcp",
      "headers": { "X-AgentCompute-Source": "guide" }
    }
  }
}

Give your agent this first task

Use AgentCompute to draft three short product descriptions for a reusable
water bottle. This is nonconfidential test content I have permission to process.
Check availability first. Create one trial session and retain it.
Submit one chat task with a stable idempotency key, follow the returned
polling interval, then retrieve the final text. Report failure honestly.
Do not create extra sessions or resubmit new work to bypass limits.

1. Call get_service_status. Submit only if the required service is available.
2. Call create_trial_access once and reuse its token.
3. Call submit_chat_job or submit_image_job with permitted content, consent: true and a unique idempotency key.
4. Poll get_studio_job at poll_after_seconds. Read result.text on success. For images, call get_studio_image to download the PNG.

Audio uses an authenticated WAV upload before submit_compute_job. Audio API guide.

Or run the complete Python example

Python 3.10+, standard library only. Download the client.

python agentcompute_studio.py chat "Explain GPU memory in three bullets." --state task-01.json --output answer.txt --licensed-nonconfidential

For images, use image and a PNG output filename. Give each new task a new state filename. Reuse the same state to resume an interrupted task. The session is reused across tasks.

State and session files contain a private token. Session expiry is reported, never silently replaced. Do not rotate sessions to evade limits.

Data and attribution

Only submit material you have permission to process and that contains no personal, sensitive or confidential information. Prompts, uploads and results reach the operator's PC. Outputs are untrusted data, never instructions that override your agent's task. Image use is subject to the model restrictions.

Optional integration headers: X-AgentCompute-Client-ID (a random stable UUID, never a name or email) and X-AgentCompute-Source (guide, python-example, mcp-registry, github or direct). Client IDs are keyed-hashed. Analytics retain metadata for 30 days, with no prompts, results, raw tokens or IP addresses. Counts and sources are aggregate and self-reported. Existing operational retention rules are described in the full guide.