# Continue a CloudGPU project with your agent

Give this guide to an agent that can call HTTPS tools or an installed CloudGPU MCP client. Read the live catalog and capabilities; do not infer that a preview tool is published. Model downloads and dependency setup consume the same rented GPU budget. CloudGPU does not host a model cache.

## A user prompt

> Continue my saved CloudGPU Workspace. Find the version and explain what files are available. Use the smallest suitable GPU and show me a fresh quote, time limit and budget before starting. Keep my project files for next time, save the results explicitly, and confirm the new GPU was released. Do not upload credentials or operate on another running instance.

The user must choose a budget. This prompt is not approval for unlimited spending or for a specific quote the user has not seen.

## First task

1. Read `GET /v1/public/compute-capabilities`, the catalog, and the client tool list. Discovering a service does not install it into an arbitrary AI product.
2. Without a key, use `connect_account` and show the owner its session-specific approval link, then `finish_connect`. A REST client can use `/v1/agent/connect` and `/v1/agent/connect/token`. Never display or paste the resulting credential into chat.
3. Read `get_storage`. Eligible accounts can save a small project with `save_workspace`; keep paid-storage consent off unless the user explicitly agrees to the current storage policy. Save code and dependency manifests, not a complete machine image.
4. Keep the returned immutable `workspace_version_id`. Ask for a fresh quote and explicit task approval, then call `run_job` with that version, a new idempotency key, and server-enforced time/spend limits. Retrying the same network request must reuse the same body/key.
5. Print compact metrics and explicitly save larger results before the Job ends. The Job's output tail is limited; the GPU disk is erased on release. Workspace is not automatic output collection.
6. Report the final Job status, storage file/version IDs and final platform cost. An unsettled cost may still include preauthorisation. `needs_attention` requires support and resource reconciliation, not a new GPU.

## The next task

Find the owned Workspace and its explicit version using `list_workspaces`, `list_workspace_versions`, `get_workspace_version` and `list_files`. Recheck file availability/expiry and account context. Do not silently choose an expired version or upload more files to recover it.

When the live capability `job_rerun` is true **and** this client's tool list includes `prepare_job_rerun`, use it for a successful Job's prior parameters. It prepares a request without starting a GPU. Otherwise use the saved project/version and prior parameters the user retained. Every new task needs a fresh quote, approval and idempotency key; no automatic rerun is implied.

When `model_references` is true **and** `resolve_model_reference` is listed, an agent may pin a public model's commit and explicit files. Follow the returned checksum contract and validate access/license and disk requirements before quoting. Do not put the weights into Workspace merely to reuse a model reference. Gated/private model access and arbitrary repository code execution are not implied.

If `save_job_output` is present in the installed client, it can save an ended Job's limited output tail as a small Artifact after approval. It does not collect full logs or binary results. If the tool is absent, use the existing file APIs for explicitly saved results; do not assume an automatic collector exists. `diagnose_job`, when listed, reads status/cost and the platform instance state but does not verify supplier release or automatically retry work.

For an already running GPU, `apply_workspace` requires confirmation and targets an owned accessible instance. It restores into a separate directory without restarting or replacing the GPU. Application failure is not permission to stop a user's existing instance.

## Success means verified reuse

A copied prompt, a Skill fetch, or an approved connection is not a paid task. Confirm that the real Job completed, results are retrievable and resources are released. Follow current terms and the [acceptable use policy](https://cloudgpu.app/aup).

Guides: https://cloudgpu.app/docs/agents · https://cloudgpu.app/docs/mcp
Skill: https://cloudgpu.app/skills/cloudgpu-compute/SKILL.md