Auxen MCP Server
This is the public manifest repo for the Auxen MCP (Model Context Protocol) server. The server itself runs at https://api.auxen.ai/mcp — this repo exists so registries (Smithery, Glama, the official MCP registry) have a canonical place to read metadata from.
Connect
The Auxen MCP server is a remote, StreamableHTTP server. Add it to your MCP client by URL:
Authentication uses OAuth 2.1 + PKCE (recommended for browser-based clients) or a direct Auxen API key (auxen_live_* / auxen_test_*) sent as Authorization: Bearer <key>.
OAuth flow
The discovery metadata is at:
https://api.auxen.ai/.well-known/oauth-authorization-server (RFC 8414)https://api.auxen.ai/.well-known/oauth-protected-resource (RFC 9728)
Clients that support Dynamic Client Registration (RFC 7591) — including Claude.ai's Connectors Directory — can register themselves automatically. After registration the client redirects the user's browser to https://api.auxen.ai/oauth/authorize, the user logs in to Auxen and approves the connection on https://auxen.ai/oauth/authorize, and the client receives an authorization code that exchanges for an access token at https://api.auxen.ai/oauth/token.
Direct key (programmatic)
For agents that don't go through a browser, generate an auxen_live_* (or auxen_test_*) key at https://auxen.ai/dashboard/api-keys and send it as Authorization: Bearer <key> on every MCP call.
| Tool | Effect | Hint |
|---|
auxen_list_models | List available models, optionally filtered by size | read-only |
auxen_get_instance_status | Get status, endpoint, api_key for an instance | read-only |
auxen_list_instances | List all instances on the account | read-only |
auxen_get_balance | Read USD credits + active subscriptions | read-only |
auxen_provision_model | Provision a new model instance — spends money | destructive |
auxen_destroy_instance | Destroy an instance — irreversible | destructive |
What is Auxen?
Auxen provisions private, dedicated GPU instances running open-source models (Llama 3.1, Qwen 2.5, Mistral, Gemma 2, Phi-3). Each instance is fully private — no shared inference, no third-party routing. Pay-per-minute billing, no subscriptions.
For the human-facing documentation:
Support