Inspiration
AI can summarize a clinical record, but causal questions require more than a confident answer. Evidence may support a hypothesis, contradict it, or suggest alternative explanations. We built CausaLens to make that reasoning explicit, traceable, and human-controlled.
What it does
CausaLens is a clinical evidence investigation workspace where an AI agent can propose a causal claim, while every claim remains grounded in the exact sentences from the patient record.
The agent can:
- search the clinical record
- retrieve a structured case summary
- create causal claims
- attach supporting and contradicting evidence
- challenge its own claims
- surface alternative explanations
- focus specific evidence
- update a claim's review status
Instead of automatically deciding what caused an outcome, CausaLens keeps the final judgment with the human: accept, reject, or leave the claim unresolved.
AI proposes. Evidence challenges. Human decides.
Why WebMCP
WebMCP turns CausaLens from a website an agent can simply view into a structured environment an agent can actually work with.
CausaLens exposes its investigation capabilities as WebMCP site tools. ChatGPT can discover these tools directly in its in-app browser and use them to investigate a case without guessing how to navigate the interface.
This enables the human and agent to share the same workspace: the agent performs evidence-intensive investigative work while the human remains responsible for consequential decisions.
How we built it
CausaLens is a web application deployed on Vercel with WebMCP integrated directly into the application.
We registered eight WebMCP tools for:
- searching the clinical record
- retrieving the case summary
- creating claims
- attaching evidence
- challenging claims
- adding alternative explanations
- updating claim status
- focusing evidence
The tools include both read and write operations. Before workspace-changing actions are executed, the agent can request human approval.
The demo uses a synthetic clinical case to demonstrate the workflow safely.
Challenges
The main challenge was designing tools that were useful to an AI agent without allowing the agent to bypass the human decision-making process.
We also had to make evidence references precise enough that claims could remain tied to specific record sentences rather than unsupported model-generated conclusions.
What we learned
WebMCP makes it possible to design interfaces for humans and agents together instead of treating the agent as a separate chatbot.
The most important design decision was not giving the agent more autonomy, but giving it structured autonomy with evidence, challenge, and human oversight built into the workflow.
What's next
CausaLens could expand beyond the synthetic case into larger longitudinal records, richer evidence provenance, multiple competing causal hypotheses, and collaborative review workflows for research and other evidence-intensive domains.
This project is a research demonstration using synthetic data and is not intended for clinical diagnosis or treatment decisions.
Built With
- ai
- chatgpt
- clinical
- human-ai-collaboration
- javascript
- next.js
- react
- typescript
- vercel
- webmcp
Log in or sign up for Devpost to join the conversation.