An integrated research environment that connects commercial, institution-hosted, and local AI models in one interface, built for biomedical discovery.
Why Biorouter
Build an agent from extensions, skills, knowledge, and the model you trust, then reuse it, schedule it, and share it. Made for sensitive data, reproducible results, and work that crosses institutions.
Bundle extensions, skills, knowledge, and a model into one reusable unit. Share it across labs to run the same task on each site's own data. The workflow carries the instructions, not the records.
Biorouter helps you organize what you learn into a connected knowledge base. The agent reads from it in every chat, so it gets more useful the more you add.
A skill is a set of instructions that teaches the agent to do one thing the same way every time. Use them to repeat a method exactly, control what the agent reads, or walk through a procedure step by step.
Extensions connect the agent to outside systems over the Model Context Protocol: clinical data warehouses, biomedical knowledge graphs, literature APIs, and your own tools. Once added, they work in every session.
Run institutional models on UCSF-hosted Azure OpenAI and Bedrock, local models with a built-in llama.cpp server (no setup) or Ollama, or commercial models like Claude, GPT, and Gemini. Pick what each task needs and switch from the same screen.
Run many agents at once across chat tabs and split panes for coding, research, and analysis. Schedule workflows to run on their own, even when the app is closed.
A guided tour
From a single question to many agents at once. Keep scrolling — the page holds still while the panels move.
Home opens with a prompt and a cached summary of your work — session and token activity, your current streak, and the chats you opened most recently — refreshed quietly in the background so you never wait on history to load.
Ask in plain language and the agent gets to work — here writing and running OMOP SQL against a UCSF clinical warehouse, reading your knowledge base, and calling skills, showing each step as it goes. Switch models mid-thread, or save a useful chat as a workflow in one click.
Every conversation is a tab, and any tab can be pulled aside into a split pane to run beside the others. Spread coding, literature review, and data analysis across as many agents as you need — each with its own history and its own model — and watch them all work at the same time.
Biorouter organizes what you learn into a connected graph. Drop in documents, paste text, or add a URL, and the agent digests them into interlinked pages it reads from in every conversation — so it gets more useful the more you add.
A workflow holds everything an agent needs in one file, so the same setup runs the same way every time. Share it across labs to keep results consistent — and to run the same task on each site's data without moving the data itself.
Extensions give the agent access to outside systems over the Model Context Protocol — clinical data warehouses, biomedical knowledge graphs, literature APIs, memory, and your own tools. Add one once and it works in every chat.
A skill teaches the agent to do one thing the same way every time. Use skills to repeat a method exactly, control what the agent reads, or walk through a procedure step by step. Each is small and shareable, and you can combine several for bigger jobs.
Local models run on a bundled llama.cpp server — no setup — or through Ollama, and never leave your machine. Institutional models run on UCSF-hosted Azure OpenAI and Bedrock for sensitive data. Commercial models include Claude, GPT, Gemini and many more. Pick what a task needs, switch whenever you want, and your API keys stay encrypted on your device.