Auxen

What auxen does

The auxen MCP server connects an MCP client to Auxen's hosted infrastructure for private model endpoints. Auxen provisions dedicated GPU instances running supported open-source models, including Llama 3.1, Qwen 2.5, Mistral, Gemma 2, and Phi-3. Instances use pay-per-minute billing rather than subscriptions, and the README describes them as private, with no shared inference or third-party routing.

The service is intended for agents that need to inspect available models and account state, then create or remove model instances. Provisioning spends money, while destroying an instance is irreversible, so clients should present those operations as consequential actions.

How auxen MCP server works

The auxen MCP server is hosted remotely at https://api.auxen.ai/mcp and communicates through StreamableHTTP. A client can authenticate through OAuth 2.1 with PKCE, which is recommended for browser-based clients, or use an Auxen API key as a bearer token. API keys use the auxen_live_* or auxen_test_* formats.

OAuth-capable clients can use Auxen's published authorization-server and protected-resource discovery documents. Clients supporting Dynamic Client Registration may register automatically, then send the user through Auxen's authorization flow before exchanging the returned code for an access token. Programmatic agents can instead create a key in the Auxen dashboard and attach it to every MCP request as Authorization: Bearer <key>.

Tools and capabilities

The six available tools divide into account inspection and instance management:

• auxen_list_models lists available models and can optionally filter by size.
• auxen_get_instance_status returns an instance's status, endpoint, and API key.
• auxen_list_instances lists instances associated with the account.
• auxen_get_balance reads USD credits and active subscriptions.
• auxen_provision_model creates a model instance and spends account funds.
• auxen_destroy_instance removes an instance permanently.

The first four tools are read-only. Provisioning and destruction change account resources and should require explicit confirmation in an agent workflow.

Setup and configuration

Add the remote MCP URL to a compatible client rather than installing a local server from this repository. For interactive clients, configure the OAuth connection using the service's discovery metadata. For non-browser agents, create an API key at https://auxen.ai/dashboard/api-keys and send it as a bearer token on each MCP call.

The manifest repository provides links to the Auxen website, architecture documentation, agent reference, LLM-oriented reference, and OpenAPI specification. These references can help determine model and API details beyond the MCP tool list.

Limitations and notes

Provisioning incurs pay-per-minute charges, and destroying an instance cannot be undone. The material does not specify model availability by GPU type, startup time, regional coverage, rate limits, or endpoint lifetime, so those details should be confirmed through Auxen's documentation or service status page. No local installation procedure or required environment variables is described; the documented deployment is the hosted StreamableHTTP endpoint.