Apppublic

hmgill/CXR-Agent

sourceHugging Faceupdated 3mo agoView on Hugging Face

1likes

App README

CXR Orchestrator — Hugging Face Docker Space

A Chainlit chat frontend for the chest-X-ray agent. Attach a CXR to a message and ask for a read; the agent triages the image, reasons over it with NV-Reason-CXR-3B (via an MCP server), localizes findings, and can author an interactive HTML overlay that opens in a right-hand dashboard panel. Optional ElevenLabs voice narrates the impression inline.

Follow-up questions in the same chat reuse the loaded image and any findings or visualizations already produced — no need to re-attach.

How it works

Per chat session the app opens the reasoning MCP connection and provisions one UnixLocalSandboxClient workspace (staging the cxr-overlay skill), keeping both alive for the session. Each turn runs one agent turn; any HTML the agent writes to the sandbox output/ directory is collected and surfaced as a selector button + the right-hand panel (plus a downloadable file). If the sandbox can't be provisioned, the app falls back to the base agent for a text-only read.

The agent is the model-driven SandboxAgent (shell + filesystem). The sandbox capabilities are ordinary function tools, so it runs on Claude via LiteLLM.

Repo layout (Space root)

Dockerfile
app.py
requirements.txt
README.md                       <- this file (HF Docker Space header)
chainlit.md
.chainlit/config.toml
public/elements/OverlayFrame.jsx
cxr_agents/  tools/  skills/  models/  sandbox/

Secrets (Space → Settings → Secrets)

Name	Required	Purpose
`ANTHROPIC_API_KEY`	yes	Claude (orchestrator, via LiteLLM)
`ELEVENLABS_API_KEY`	no	Voice narration (a free key works)
`CXR_REASON_MCP_URL`	no	Override the reasoning MCP URL
`ELEVENLABS_VOICE_ID`	no	Default voice (name or raw ID)

Disclaimer

Decision support only — not a diagnostic device. Always confirm findings with a qualified radiologist.