Skip to main content
Valkyrie ships a built-in Model Context Protocol (MCP) server. Connect it to an MCP-capable client, Claude Code, Claude Desktop, Cursor, and others, and the assistant gets tools to operate your Valkyrie account: deploy and manage models, run and monitor fine-tuning jobs, inspect GPUs and cloud state, read logs, and work with datasets.
This is the control plane, it’s how an agent operates Valkyrie. It is not how you chat with a model. To use a Valkyrie-hosted model as your LLM, use the inference API instead, see Connect a model. The two are separate and can be used together (e.g. a Valkyrie model as your agent’s brain, plus these tools to manage infrastructure).

​
What it can do

Deployments

List, deploy (base or fine-tuned), start/stop/restart/resume/pause, recover, health-check, and schedule model deployments.

Fine-tuning

Create, list, monitor, retry, cancel, and recover jobs; list tuned models; fetch and download outputs.

Datasets

List and inspect datasets, get an upload URL, download, and delete.

Cloud & GPUs

Check account status/balances, GPU availability, and running instances; sync, adopt, or stop instances.

Logs

Read and collect job, deployment, and vLLM service logs.

Inference pools

Inspect and manage pool configuration, capacity, and member leases.
The full, explained list is in MCP tools.

​
What it can’t do (and other limits)

Read these before you connect a key.
  • It is not the inference endpoint. The MCP server operates infrastructure; it doesn’t replace the OpenAI/Anthropic-compatible API you call to actually run prompts against a model.
  • There is no read-only mode. The tool set always includes mutating tools, deploy_base_model, stop_deployment, terminate_cluster, delete_dataset, and more. They act on whatever environment your key points at, and some spend from your wallet or tear down running resources. There is no dry-run or undo for a terminate.
  • Keys are per-environment. A key operates only its own environment; a key from one environment returns Invalid API key against another. Cross-environment ownership is guarded, you can’t operate another environment’s resources.
  • Logs are workload logs, not platform logs. The log tools return your job, deployment, and serving-engine output, not Valkyrie’s own internal application or crash logs.
  • Pool administration never provisions or mutates serving deployments. Inference-pool tools manage pool configuration and leases only; they don’t create or change the underlying model deployments.
  • It doesn’t manage account administration. Users, roles, billing top-ups, and dashboard account settings are handled in the dashboard, not through these tools.

​
Next step

Set up the MCP server

Register Valkyrie in your client, with the transport and auth details that matter.