Resource
MindRouter
Most campus AI discussions start with which model to buy and stop before the harder question: where it runs, who meters it, and who can see the logs. MindRouter is an answer to that second question. It is a self-hosted gateway that sits in front of a heterogeneous cluster of LLM servers (vLLM and Ollama nodes on GPUs the institution owns) and exposes one API that speaks the OpenAI, Anthropic and Ollama dialects, translating between them so existing client code keeps working.
What you can do with it
Point a tool at the gateway instead of at a vendor endpoint. Behind it, MindRouter handles the parts a shared deployment needs: fair-share scheduling with weighted deficit round robin, per-user token and request quotas, role-based access with optional SSO through Azure AD/Entra, Google, any OIDC provider or SAML 2.0, and full audit logging of prompts, responses and artifacts. A GPU sidecar agent reports per-node utilization, memory, temperature and power, so an administrator can see energy cost per request and drain a backend gracefully when maintenance is due. It covers more than chat: embeddings, vision, image and video generation, text-to-speech and speech-to-text, document OCR. It also ships MCP servers and agent skills for coding assistants.
The case for a college is the one the project names: institutional AI sovereignty. Prompts and documents stay on infrastructure the institution operates rather than passing through a third-party cloud, which is easier to defend to a faculty senate or a privacy office than a data processing agreement with a vendor. Whether that argument holds depends on whether the institution has the GPUs, the power and the staff. The gateway makes local hosting manageable, not cheap.
Who it is for
Research computing and central IT groups with existing GPU capacity, teaching and learning centers that want a single place to meter faculty and student use, and administrators weighing hosted enterprise platforms against running inference themselves. It is a Docker Compose deployment with documented SSO configuration and custom branding, and it is maintained at the University of Idaho, where the author presented it at PEARC '26; the project has 22 stars and 3 forks as of November 2026.
Notes and caveats
Free and open source under Apache-2.0, but the cost moves rather than disappears: hardware, electricity, and the staff time to run a cluster. The software is young by infrastructure standards, with three contributors, a single originating institution and releases numbered in the hundreds of tags, so a college should pilot it beside an existing service rather than as the only path to a model. Published evaluation of the scheduling and quota behavior under real campus load is not in the repository; the PEARC paper is the place to look for the authors' own account.
Connected Concepts
Connected Resources
- Amplify GenAIVanderbilt's MIT-licensed enterprise generative AI platform, deployed in an institution's own AWS account: custom assistants over local documents, retrieval-augmented answers, sharing across the organization, and cost accounting down to the individual message.