Integration Guide
Connect Vast.ai to Hermes Agent or OpenClaw with Composio
Connect your Vast account, verify your balance and instance list, then ask your agent to compare GPU offers or manage a rental. The dashboard steps are the same for both managed frameworks.
What this connection gives your agent
Vast.ai supplies the GPU machines. Composio stores the connection authorization. OpenClaw Launch makes the supported operations available to your managed Hermes Agent or OpenClaw. You keep using your own Vast account and its balance.
As of September 14, 2026, the managed integration exposes 44 tools: 41 native Vast operations added by OpenClaw Launch alongside the three searches in Composio's standard Vast toolkit. A Connected badge means the connection was saved. A successful account query is the next check; neither signal proves that every write operation is permitted.
1. Create a Vast API key
Have a Vast account and a running managed agent ready. If Vast is already connected, go straight to step 3 and test the saved authorization before creating another key.
- Sign in to the Vast console and open Keys.
- Under API Keys, click +New and give the key a recognizable name, such as
my-managed-agent. - Select the permissions needed for your intended tasks. Review the defaults: Vast keys can start with full account access.
- Create the key and copy it into the connection form in step 2. Store a backup in your password manager.
Use a durable API key, rather than an SSH key or temporary session key. Follow Vast's key-creation instructions if the console layout has changed. Keep the key out of bot messages and screenshots.
Which permissions are needed?
| Task | Permission category to check |
|---|---|
| Find offers and benchmarks | misc |
| Read account balance | user_read |
| Read instances and logs | instance_read |
| Create, start, stop, or delete instances | instance_write |
| Read billing history | billing_read |
| Manage account SSH public keys | user_write |
These are categories from the Vast permissions reference; resource constraints and individual endpoints can narrow access further. Enable the tasks you need. Connecting through Composio cannot add permissions to your Vast key.
2. Connect Vast.ai in the dashboard
- Open Dashboard → Integrations.
- Search for Vast and select Connect on the Vast.ai card.
- Enter your Vast API key in the API key field of the connection form and complete Connect.
- Wait for the card to show Connected, then return to your agent.
The form sends the credential to Composio to establish your connection. Use this same flow for managed Hermes Agent and OpenClaw; you do not need to install a separate Vast CLI or configure an MCP server for the managed route. A Composio project key is a different credential and does not belong in the Vast API key field.
3. Verify access without creating a rental
Ask your agent:
Use my connected Vast.ai account. Refresh the available Vast tools, read my balance, and list my existing instances. Report whether each query succeeds. Do not create, start, stop, or delete anything.
A successful balance response and instance list verify these reads. An empty instance list is normal if you have no rentals. A successful public offer search alone does not verify access to your account or prove permission to rent a machine.
Commands the managed agent can use
These commands run inside your managed agent's terminal. You can ask the bot to run them; there is no API key to paste into the command.
composio connections
composio search vast_ai
composio execute VAST_AI_OCL_GET_ACCOUNT '{}'
composio execute VAST_AI_OCL_LIST_INSTANCES '{}'If multiple Vast connections exist, select the intended account explicitly. The agent should inspect the current action schema and connection list instead of guessing which account to use.
4. Compare the full cost before renting
Give the agent the GPU, disk allocation, expected runtime, and data-transfer needs. For example:
Find available single RTX 3090 offers with 20 GB of allocated storage. Use the full-pricing search. Show the offer ID, GPU hourly price, storage cost, upload and download rates, and an estimate for two hours. Mark unknown costs clearly and wait for me to choose an offer before renting.
For the pricing query, the managed agent can run:
composio execute VAST_AI_OCL_SEARCH_GPU_OFFERS_FULL '{"limit":5,"allocated_storage":20,"filters":{"gpu_name":{"eq":"RTX 3090"},"num_gpus":{"eq":1}}}'Use your intended disk size in allocated_storage. The native response preserves dph_total, storage rates, and network rates when Vast returns them. Read the price breakdown before adding components: storage may already be included in the quoted total for that allocation. Missing fields mean unknown costs, not free usage. Transfer charges depend on how much data moves. See Vast's pricing guide.
Once you choose an offer, specify the disk size, a versioned container image or template, and your spending limit. Ask the agent to summarize the rental before you authorize creation. Afterward, have it read the new instance and verify the actual status. API acceptance does not mean the machine is ready for SSH or that your application has started.
Available functions
| Area | Supported managed operations |
|---|---|
| Account and pricing | Balance, invoices, detailed charges, full GPU and volume offers, templates, and benchmark results. |
| GPU instances | List, inspect, create, start, stop, destroy, reboot, recycle, label, change an interruptible bid, read deposit details, and prepay. |
| Persistent volumes | List, create, clone, and delete volumes. |
| SSH access | List, add, update, and delete public keys; list instance keys; attach and detach instance access. |
| Data transfers | Copy between instances, cancel copies, list existing cloud connections, and start or cancel cloud copies. |
| Diagnostics | Request instance logs and queue commands supported by Vast's constrained command API. |
| Serverless resources | Read endpoints, worker groups, deployments, and individual deployment status. |
Resource changes require your explicit request for the particular operation and any cost. Log and command requests can return a queued result URL; the agent must check the result before claiming completion. This integration does not expose API-key administration, payment-method changes, credit transfers, team or host administration, or serverless deployment changes.
Self-hosted agents: a separate Composio setup
If you run the agent yourself, connect Vast in your own Composio account and configure a compatible client. The generic routes below expose the tools provided by your Composio setup. They do not install OpenClaw Launch's 41 additional VAST_AI_OCL_ operations. Native Vast controls outside the managed platform require your own API or CLI integration.
Hermes Agent
Composio documents an HTTP MCP connection for Hermes. Merge this into ~/.hermes/config.yaml, replacing the placeholder with the Composio credential for this MCP service. The header takes a Composio key, not your Vast key.
mcp_servers:
composio:
url: "https://connect.composio.dev/mcp"
headers:
x-consumer-api-key: "YOUR_COMPOSIO_API_KEY"
connect_timeout: 60
timeout: 180Keep the configuration private. Start a new Hermes session, or use /reload-mcp in an active CLI session. Follow the authentication and app-linking steps in Composio's Hermes setup guide and the Hermes MCP reference.
OpenClaw
For a self-hosted OpenClaw installation, install Composio's skill for configuration guidance:
npx skills add ComposioHQ/composio --skill composio --agent openclaw --globalStart a new OpenClaw session so it can load the skill. To connect your own accounts for app actions, open Composio Connect, select OpenClaw under More clients, and follow that client's setup instructions. Link Vast for the intended account and verify the tools that become available. Installing the skill alone does not establish the connection. See the official Composio client instructions and our OpenClaw Composio guide.
Troubleshooting
| Problem | Next step |
|---|---|
| Connected, but the agent says controls are missing | Ask it to run composio search vast_ai and inspect the current catalog. On managed instances, the expanded actions have the VAST_AI_OCL_ prefix. |
| The quote omits disk or network costs | Use VAST_AI_OCL_SEARCH_GPU_OFFERS_FULL with the intended disk allocation. Do not use the simplified search response as a complete budget. |
| Balance is unavailable | Use VAST_AI_OCL_GET_ACCOUNT. Check the actual result before concluding that the key lacks permission or asking for a replacement. |
| Vast returns 401 or 403 | Check the error and key status. A revoked or replaced key needs reconnecting; a restricted key needs the permissions for the requested operation. Read access does not imply write access. |
| More than one active Vast connection | Select the intended connected account. Do not automatically use a different account because a request failed. |
| A create, payment, or other change times out | Check resource state and balance before retrying. An uncertain response can mean the operation already happened. |
| The instance stays in Scheduling | A stopped instance may have to wait for GPU availability. Inspect its status before deciding whether to wait or choose another offer. |
| API controls work but SSH fails | Check the instance's running status, SSH host and port, and authorized public key. The client making the connection needs the matching private key. |
Vast.ai connection FAQ
Does the Vast.ai connection work on Hermes Agent?
Yes. Managed Hermes Agent and OpenClaw use the same OpenClaw Launch Integrations page and Vast tools. Connect the account in the dashboard, then ask your agent to verify access. Self-hosted agents need their own integration setup.
Do existing users need to reconnect to get the expanded controls?
A valid saved Vast connection is reused. Existing and future connections on OpenClaw Launch receive the expanded tool catalog without a container restart. Each API key must still allow the requested operation. Reconnect when the saved credential is invalid or has been replaced.
Does connecting Vast rent a GPU or start billing?
Connecting and reading offers do not create a rental. Creating a GPU instance or volume can incur Vast charges. Vast bills GPU, storage, and network use separately from your OpenClaw Launch subscription; stopping an instance leaves storage charges running.
Why does the agent only see three Vast search tools?
The standard Composio Vast toolkit lists GPU offers, volume offers, and templates. OpenClaw Launch adds native Vast operations through its managed integration. Ask a managed agent to refresh discovery with composio search vast_ai. A generic self-hosted Composio MCP server does not automatically include the platform additions.
Can the agent run Blender or download my rendered files?
The integration can manage a GPU instance, its SSH public keys, transfers supported by Vast, logs, and constrained commands. Installing applications, running a full rendering workflow, and transferring files to your own computer generally need a separate SSH or scp connection. Connecting the API alone does not establish that connection.