Give your AI agent access to GPU compute

CloudGPU lets an AI agent find a GPU, see what it costs, and run work on it within a budget you set. Discovery and quotes need no account. Running anything needs an API key, and CloudGPU’s servers enforce the budget and time limit, so a GPU still shuts down if the agent’s session ends. You can give an agent its own key with a spending limit through a connect link, without copying a key into it.

Agent docs · MCP server · OpenAPI · Live prices

GPU compute for AI agents: quick answers

Question	Current CloudGPU capability	Reference
Can an agent inspect GPUs before signing in?	The GPU catalog and quotes are public; starting paid work requires approved access	Public catalog
Can a task have a spending limit?	Jobs take explicit GPU spending and runtime limits; storage fees require separate approval	Jobs API
Can project files be reused?	Save an immutable Workspace version, restore it before a Job or confirm application to an existing GPU	Workspace and MCP
Can selected results survive GPU release?	Request output persistence explicitly and verify that collection succeeded	Practical Job guide
Is there a shared public model cache?	Public pinned model files download to the rented GPU; this is not a shared cache	Model preparation
Can several Jobs be submitted together?	MCP supports a bounded 1–8 Job batch with partial receipts; not automatic retries of failed executions	Batch limits

These facts were checked against the published API and MCP 0.11.0 on 2026-11-11. Read live capabilities before relying on an optional feature. Availability and prices change; get a fresh quote.

Connect your agent with an owner-approved link

Copy the instructions above to an agent that has HTTPS tools or a configured CloudGPU MCP client. The agent reads the compute Skill, discovers capabilities and obtains a quote. It then generates a unique approval link for its current session. You sign in and confirm the agent, budget, concurrency and expiry; the agent collects its own scoped key. No secret needs to be copied into chat.

Opening this page or generating an approval link does not start a GPU. Connection approval is not unlimited permission to spend: the actual task, quote and limits still require your explicit approval. Clients without tool access need to be configured first; this is not a universal one-click installation into every AI product.

How it works

Step	What the agent calls	Needs a key?
Discover	GET /v1/public/gpu-catalog: GPU models, VRAM, live price per GPU-hour, availability, billing rules, system disk options, templates	No
Quote	POST /v1/quotes with VRAM, GPU count, template, duration and budget: candidates sorted by estimated cost	No (send a key to also see whether your balance can start each one)
Connect	POST /v1/agent/connect: the account owner opens a link and approves the agent with a spending limit	No (this is how the agent gets one)
Run a job	POST /v1/jobs with a command, max_minutes, max_spend_usd and an Idempotency-Key: CloudGPU starts a GPU, runs the command and destroys the GPU	Yes
Or rent	POST /v1/gpu/instances, then SSH or the MCP server’s upload, run_command and download tools, then POST /v1/gpu/instances/{id}/destroy	Yes

What exists today

• Public GPU catalog at https://cloudgpu.app/v1/public/gpu-catalog, JSON, with generated_at and cache_ttl_seconds. It uses the same prices as the website.
• Quote API at POST https://cloudgpu.app/v1/quotes. Each candidate has estimated_total_cost (what the task is expected to use) and required_starting_balance (what your balance needs to start it). These differ because one hour is pre-authorised at start. Quotes are estimates and expire after 5 minutes.
• Jobs API at /v1/jobs: one command, run start to finish on CloudGPU’s servers. Retrying with the same Idempotency-Key returns the same job and never starts a second GPU. Each job has a status, an event history, its exit code, the last part of its output and its cost.
• Agent connect: an agent asks for a key with POST /v1/agent/connect; you open the link, sign in, check the agent’s name and limits, and approve. The key has a total spending limit, a number of GPUs it may run at once and an expiry date. Revoke it on the API page and it can no longer start anything.
• GPU instances API under /v1/gpu/: list offers and templates, start an instance with server-enforced limits (max_minutes, max_spend_usd, idle_minutes), an optional startup script and an optional larger system disk, read its status and SSH details, destroy it.
• MCP server: npx -y cloudgpu-mcp, for Cursor, Cline, Claude Code, opencode and other MCP clients. It runs on your computer and drives the same API, jobs and connect links included (guide).
• Workspace and persistent files: explicitly save project files and immutable version manifests, restore them before a Job runs, or apply a saved version to your running GPU without restarting it. Open Workspace; use authenticated GET /v1/storage for eligibility, quota, retention and charges. The MCP tools expose the same file and Workspace features.
• Model API: the same key works for the OpenAI-compatible model API at https://cloudgpu.app/v1 (docs).

Destroying an instance, or the end of a Job, erases its local disk. Eligible funded accounts can explicitly save project files in Workspace independently of the GPU, restore a saved version in a later Job, or apply it to a running GPU. Saved files remain subject to storage quotas and retention rules. Workspace does not automatically back up the disk. Jobs can explicitly collect selected outputs with persist_results; confirm collection succeeded and check the applicable storage eligibility, quota, retention and charges. Unsaved files are not preserved.

Billing and limits

Per minute, rounded up, with a 1-minute minimum. One hour is pre-authorised while an instance runs, and unused minutes go back to your balance when you destroy the instance. There is no subscription and no minimum term.

Billing comes from a prepaid USD balance. Every job needs a time limit and a spend limit, and every instance started through the API can carry them; CloudGPU’s servers enforce both. A key from a connect link also has a total spending limit and a cap on how many jobs or instances it runs at once.

Example: quote, then run a job

# 1. What would 30 minutes on a 24 GB card cost? (no key needed)
curl -s https://cloudgpu.app/v1/quotes \
  -H 'Content-Type: application/json' \
  -d '{"requirements":{"min_vram_gb":24},"estimated_duration_seconds":1800,"max_budget_usd":"1.00"}'

# 2. Save this key once; reuse it with the same request after a lost response
IDEMPOTENCY_KEY=$(uuidgen)
# Run a job on the first candidate with a hard time and spend limit
curl -s https://cloudgpu.app/v1/jobs \
  -H "Authorization: Bearer $CLOUDGPU_API_KEY" -H 'Content-Type: application/json' \
  -H "Idempotency-Key: $IDEMPOTENCY_KEY" \
  -d '{"quote_id":"quote_...","command":"python train.py && rclone copy out/ mystorage:run1","max_minutes":40,"max_spend_usd":1.00}'

# 3. Check it
curl -s https://cloudgpu.app/v1/jobs/job_... -H "Authorization: Bearer $CLOUDGPU_API_KEY"

If the quote has expired, the GPU is no longer available, its price went up or the configuration differs from the quote, step 2 returns an error and starts nothing. If step 2 times out, send it again with the same Idempotency-Key.

FAQ

How do I give an agent a key without giving it my whole account?

Let the agent call POST /v1/agent/connect (the MCP server’s connect_account tool does this). You get a link; sign in, check the agent’s name, the spending limit and how many GPUs it may run at once, lower them if you like, and approve. The agent receives its own key. CloudGPU’s servers refuse anything over those limits, and you can revoke the key on the API page at any time.

What if the agent sends the same job twice?

Jobs need an Idempotency-Key. The same key with the same request returns the job that already exists, so a retry after a timeout never starts a second GPU. The same key with different parameters is refused.

Does an agent need an account to see prices?

No. The GPU catalog and the quote API are public. An API key is only needed to start an instance, and with a key the quote also says whether your balance can start each candidate.

How do I stop an agent from overspending?

Set max_minutes and max_spend_usd when the agent starts an instance. CloudGPU’s servers destroy the instance when either is reached, even if the agent’s session has ended. The MCP server also takes a per-session spend cap.

Why is the balance needed to start higher than the quote?

One hour is pre-authorised while an instance runs. A 30-minute task may be quoted below the hourly price, but your balance must cover the hour to start it. Unused minutes go back to your balance when you destroy the instance.

Which agents can use CloudGPU?

Any agent that can make HTTPS requests can use the REST API. MCP clients such as Cursor, Cline, Claude Code and opencode can use the MCP server.

Does CloudGPU keep my files after the instance is destroyed?

Destroying an instance, or the end of a Job, erases its local disk. Eligible funded accounts can explicitly save project files in Workspace independently of the GPU, restore a saved version in a later Job, or apply it to a running GPU. Saved files remain subject to storage quotas and retention rules. Workspace does not automatically back up the disk. Jobs can explicitly collect selected outputs with persist_results; confirm collection succeeded and check the applicable storage eligibility, quota, retention and charges. Unsaved files are not preserved.

Can ChatGPT search access and tools execute the same actions?

No. Reading or citing a CloudGPU page does not authorize a rental or install tools. Executing a task requires a compatible client with configured CloudGPU MCP tools or HTTPS access, an owner-approved credential and an explicitly approved quote. The published MCP package uses local stdio; an account approval link alone does not configure an arbitrary ChatGPT chat.

Will ending the chat stop a submitted Job?

Not by itself. An accepted Job can continue on the backend independently of the client. Spending and runtime limits still apply. Check the Job status and release evidence; do not infer release from a disconnected chat or treat stopping or needs_attention as confirmed release.

Are public models already cached on every GPU?

No. Supported public model references pin a repository revision and file checksums for download onto the rented GPU. A project version does not preserve installed dependencies, and download time is part of the rental cost. Shared model caching is not included.

---

Prices are on https://cloudgpu.app/pricing and in the public catalog, and are not repeated here.

Continue your project through an agent

Tell your agent: “Continue my saved CloudGPU Workspace. Check the saved version and available files, then show me a fresh quote before starting.” The agent can reuse project files without requiring a separate task console. Every new task still needs a quote and your approval; saved files have their own quota and expiry.

Use the project reuse example with an HTTPS-capable agent or configured MCP client. It explains first-run authorization, Workspace reuse, result handling and safe recovery. Read live capabilities and the installed tool list before using optional features; a copied prompt does not install tools or start a GPU. Models and installed dependencies are not automatically preserved with project files.