POC — MCP Server + Clinical Agent

POC of technical feasibility for the two services in the Phase 2 plan, built on a custom Model Context Protocol server and the Anthropic API.

1. Document scanning and analysis — triage, read, structure, suggest renewal for the doctor to approve, and investigate with the patient what the documents did not answer.

2. Anamnesis agent — deduce the script from the doctor's description, validate it, register it as a specialty skill, and route from the general practitioner to the specialist when the case requires it.

There are 11 flows / 70 steps, executed against the real MCP server. No response was simulated: where a credential was missing, the step stopped.

Verdict, in two lines

Goal 1 — usable, with one caveat that the POC alone does not resolve: accuracy on real handwriting was not measured, and that is what decides whether the service is useful. Goal 2 — the mechanism is in place; whether an LLM-deduced script is clinically good is a medical judgment on content, not technical feasibility.

The full verdict, with what was validated step by step, is in findings.md.

Related MCP server: atlas_mcp

How to run

cp .env.example .env       # preencha ANTHROPIC_API_KEY
npm install
npm run poc                # servidor MCP (5190) + viewer (5180)

Open the viewer at http://localhost:5180 and execute the steps.

Before the first step: there is no sandbox — every execution is a billing event. Set a spend limit in the Anthropic console. MANUAL-SETUP.md has the step-by-step for this and everything else only you can do.

Suggested order: doc-triagem/step-01 first (the chair, ~US$ 0.002, the cheapest step — if it fails, nothing after is worth it), then the happy doc-* flows, doc-lote, the ana-* ones, and finally the alternative ones. Running everything costs about US$ 0.60; doc-lote is the most expensive (~US$ 0.22 of the US$ 0.60).

npm run check    # tsc + validação dos fluxos + suíte de defesas (27 casos)

The viewer

The viewer is not decoration — it is where the POC is read. Each step shows:

• the raw HTTP request that goes to the MCP server, editable field by field;

• the fixture it consumes, rendered inline (photos as images, PDFs in a frame) — and with a Replace… that accepts a photo of yours, handwritten, to re-execute without touching any file;

• the code trail the call goes through, in order: transport → tool → domain → LLM. Clickable, opens in the editor (set POC_EDITOR in .env).

How it is set up

MCP server in Streamable HTTP stateless — sessionIdGenerator: undefined, enableJsonResponse: true, new server and transport instance per request. Each call becomes a POST JSON-RPC with JSON response, which is what allows the viewer to show the call as the HTTP it is. Node 22 running .ts directly via tsx, no build.

mcp-server/      servidor MCP: transporte, tools, domínio, chamadas à LLM
handlers/        os poucos passos cujo corpo não cabe num template declarativo
flows/           os 11 fluxos, em steps.json editáveis
fixtures/        receitas, exames, PDFs — inclusive os de prompt injection
public/          o viewer
server.ts        o viewer (back-end)
checar-defesas.ts / checar-fluxos.py    a suíte que `npm run check` roda

Two things cut across both goals and stand on their own:

The three-stage pipeline paid for itself twice. Cost: triage runs on Haiku and answers a cheap perceptual question before OCR on Sonnet looks at the image. Safety: separating reading from interpretation gave a natural place for injection to be transcribed, flagged, and refused instead of obeyed.

The deterministic rule is what prevents the suggestion from becoming a decision. Renewal vetoes, spending cap, skill catalog, doctor validation — none depend on the model remembering on its own, and none costs a token.

Scope

This POC covered the simplest use case that works. Security, performance, and regulatory are deliberately out of scope and are work for the next steps — findings.md names each one in "Open issues".

Models retire on a set date — that is why ANTHROPIC_MODEL and ANTHROPIC_MODEL_TRIAGEM live in .env and not in code. findings.md records the known dates at the time of execution.

Documents