Rent a Cloud GPU with Hermes Agent or OpenClaw

What this looks like in practice

We tested this on our own managed Hermes Agent while it was producing an animated film in Blender. It had been rendering on CPU and needed a full-HD, 128-sample Cycles frame to check quality. From one chat message, the bot:

1. searched Vast.ai for GPUs with at least 12 GB of VRAM, enough CPU and disk, and a CUDA driver new enough for its Blender build, all under a price cap;
2. rented an RTX 4060 Ti 16GB with a 20 GB disk and a throwaway SSH key it generated for this one job;
3. uploaded Blender and the scene, checking a SHA-256 hash for each upload;
4. confirmed Cycles was really rendering on the GPU (CUDA) and not falling back to CPU;
5. rendered the 1920×1080 frame in 22.5 seconds, then downloaded it and checked the file;
6. destroyed the instance, removed the SSH key, and confirmed nothing was left running.

The whole rental, from creating the machine to releasing it, took 194 seconds. Vast posted a $0.007 charge for that instance. The bot then compared the GPU frame with its earlier CPU preview and told us honestly what had improved and what still fell short, rather than declaring the job done.

How it works

What you need

• A running managed Hermes Agent or OpenClaw bot on OpenClaw Launch.
• A Vast.ai account with credit or a payment method on file.
• Vast.ai connected to your bot. Follow the Vast.ai connection guide: create a Vast API key with the permissions this workflow uses (misc for offer search, user_read and user_write for balance and SSH keys, instance_read and instance_write for rentals, billing_read for charges), then connect it under Dashboard → Integrations. That takes about two minutes.

Before your first rental, check the connection without spending anything:

Use my connected Vast.ai account. Read my balance and list my instances. Do not create, start, stop or delete anything.

1. Describe the job and your limits

Tell the bot what the job is, what counts as success, and how much you are willing to spend. The more concrete the limits, the better the quote. For example:

Render frame 289 of my Blender project at 1920×1080 with 128 Cycles samples on a rented GPU. Use a single GPU with at least 12 GB of VRAM, a 20 GB disk, and a maximum of $0.40 per hour. Quote the full cost first and wait for my approval. After the render, download the frame, verify it, destroy the instance, and remove any SSH key you added.

If you do not know how much VRAM your job needs, ask the bot to work it out from your model or scene first. Specifying VRAM, CUDA version and disk up front removes most of the cheap offers that would fail halfway.

2. Approve the quote

The bot searches current Vast offers with full pricing and comes back with a specific machine. A good quote names:

• the offer ID, GPU model, VRAM, and the host's maximum supported CUDA version;
• the hourly price for the disk size you asked for (Vast includes storage in that quote, so it should not be added twice);
• upload and download rates, and an estimate of how much data will move;
• an estimated total for the expected runtime, and a hard time limit after which it will release the machine.

Offers on Vast change minute by minute. If the chosen offer disappears, the bot should come back with a new quote instead of quietly renting a different machine. Reply with a clear approval, for example "Approved: offer 123456, up to $0.40 per hour, 30 minutes max."

3. The bot rents the machine and connects

Once you approve, the bot typically:

1. generates a new SSH key pair in its own workspace for this job only, and registers the public key with Vast (the private key never leaves the bot);
2. creates the instance from the approved offer with a versioned Docker image; the connection refuses latest or untagged images, so the machine is reproducible;
3. checks the instance until SSH actually answers. A "running" status from Vast does not yet mean the machine is ready to connect to.

The rental calls look like this from the bot's terminal. You do not need to type them; they show what the bot is doing on your behalf:

composio execute VAST_AI_OCL_SEARCH_GPU_OFFERS_FULL '{"limit":20,"allocated_storage":20,"filters":{"gpu_ram":{"gte":12000},"num_gpus":{"eq":1},"dph_total":{"lte":0.4}}}'
composio execute VAST_AI_OCL_ADD_SSH_KEY '{"user_confirmed":true,"ssh_key":"<the job's public key>"}'
composio execute VAST_AI_OCL_CREATE_INSTANCE '{"user_confirmed":true,"offer_id":123456,"image":"nvidia/cuda:12.8.0-runtime-ubuntu24.04","disk":20,"runtype":"ssh"}'
composio execute VAST_AI_OCL_GET_INSTANCE '{"instance_id":52500115}'

4. The bot runs the job

Over SSH, the bot uploads its inputs, installs or unpacks the software, and runs the job. Two checks separate a real GPU run from an expensive CPU run:

• Verify the uploads. Comparing a hash of each file on both ends catches a truncated upload before you pay for a broken run.
• Prove the GPU is used. Run nvidia-smi, and for Blender check that Cycles has the CUDA or OptiX device enabled. Many tools silently fall back to CPU when the driver and the software's CUDA version do not match.

5. Download, verify, release

When the job finishes, the bot copies the outputs back into its own workspace with scp and checks them. Only then does it destroy the instance:

composio execute VAST_AI_OCL_DESTROY_INSTANCE '{"user_confirmed":true,"instance_id":52500115}'
composio execute VAST_AI_OCL_LIST_INSTANCES '{}'
composio execute VAST_AI_OCL_DELETE_SSH_KEY '{"user_confirmed":true,"ssh_key_id":98765}'

The job is only finished when the instance list is empty and the job's SSH key is gone. The results then sit in your bot's workspace, where it can send them to you in chat, publish them, or keep working on them.

Jobs worth sending to a GPU

Keep costs under control

• Give a price cap and a time limit in the first message, and ask the bot to release the machine automatically when either is reached, including on failure.
• Test small first. One frame or a short sample on the GPU tells you the real speed and cost before you pay for the full batch.
• Check the actual charge. Ask the bot to read your Vast charges for that instance ID afterwards. Vast posts charges with a delay, so the first figure is not always the final one.
• Never retry a rental blindly. If a create call times out, the machine may already exist. The bot should list your instances before trying again, so you do not end up paying for two.

Hermes Agent and OpenClaw: same steps

On OpenClaw Launch, managed Hermes Agent and OpenClaw bots get the same Vast.ai connection, the same composio command and the same preinstalled ssh, scp and ssh-keygen in their workspace. Every prompt and command on this page works unchanged on both. Hermes Agent can also save the finished flow as a reusable skill, so the next GPU job starts from a tested routine instead of from scratch.

Self-hosted agents

The VAST_AI_OCL_ actions are part of OpenClaw Launch's managed connection, so a self-hosted Hermes Agent or OpenClaw does not have them. The same workflow works with Vast's own command-line tool, installed where your agent runs and authorized with your Vast API key:

pip install vastai
vastai set api-key YOUR_VAST_API_KEY
vastai search offers 'gpu_ram>=12 num_gpus=1 rentable=true'
vastai create instance OFFER_ID --image nvidia/cuda:12.8.0-runtime-ubuntu24.04 --disk 20 --ssh
vastai show instances
vastai destroy instance INSTANCE_ID

On a self-hosted agent, vastai set api-key saves the key to a file on that machine and the agent can spend with it, so give it a restricted key and put the same approval and cleanup rules in your instructions.

Troubleshooting

Cloud GPU FAQ

Related guides

• Connect Vast.ai to Hermes & OpenClaw
• Blender MCP server
• Hermes Agent + ComfyUI
• Hermes Agent + vLLM
• Faster Whisper transcription
• Managed Hermes Agent hosting