Free Claude Desktop extension · local file analysis
Is this feed ready for a real workflow?
Ask Claude to inspect a healthcare event feed. Get measured delivery delays, duplicate IDs, missing fields, and FHIR Encounter issues with row numbers and suggested next checks.
A concrete result
What the included sample finds
A six-row synthetic discharge and admission CSV gives Claude a specific investigation to work through. See the one-page visual overview.
At a 24-hour target, one event arrived after 31 hours. One ID is duplicated. Another row has a receipt time earlier than its event time. Change the target to 48 hours and the late count becomes zero.
Claude can explain the tradeoff and turn these findings into questions for the source and ingestion owners. The extension reports the evidence; it does not infer missing patients or clinical outcomes.
Install and try
From download to a useful question
- Claude Desktop: Download the
.mcpbfile above. In Claude Desktop, open Settings → Extensions → Advanced settings → Install Extension… and select it. Restart Claude Desktop if the tool does not appear. - Cursor, Windsurf, or Claude CLI: Install via Smithery:
npx -y @smithery/cli install @apratsunrthd/refhealth-feed-triage --client claude(or--client cursor). - Ask: “Use Healthcare Feed Triage on its included sample. Could this feed support a 24-hour follow-up trigger? Show the evidence and what I should ask the data owner.”
- To inspect your own file, give Claude its absolute local path and a delivery target. Example: “Inspect /path/to/events.csv against 48 hours. Which rows need investigation?”
The extension includes a synthetic CSV and a small synthetic FHIR Bundle. It requires no account or API key. Claude Desktop includes the Node.js runtime needed to run the extension. Review the source code, Smithery config, and tests on GitHub.
Supported inputs
Start with one feed
CSV event feed: columns event_id, patient_id, event_type, event_time, and received_at. Timestamps need an explicit timezone. The tool measures event-to-receipt delay, counts events by type, and flags duplicate IDs, invalid rows, and late delivery.
FHIR Bundle JSON: checks Encounter IDs, patient references within the supplied Bundle, and encounter period end times. A FHIR Bundle normally has no feed receipt timestamp, so the tool does not claim to measure delivery latency from meta.lastUpdated.
Files are limited to 10 MB and 20,000 rows or entries. The extension returns aggregate counts and row-numbered findings to Claude, not raw records or patient IDs.
After the check
Turn a finding into a source question
- Late delivery: compare event, publish, and receipt timestamps. Check timezone offsets, retries, and queue delays with the source and ingestion owners.
- Duplicates or invalid rows: confirm the authoritative event ID and retry behavior. Repair missing fields or clock errors at the source.
- FHIR references: check whether a Patient is supplied outside this Bundle before treating an unresolved in-file reference as broken. Verify Encounter end times with the source.
Use the row or entry numbers in Claude's result to locate records locally. Before using a feed in care operations, separately verify coverage, consent, clinical policy, and downstream delivery.
Data boundary
Use synthetic or approved de-identified data.
The extension reads a file on your machine and makes no network call to ref(health). Its findings are sent to Claude for interpretation. Do not use it with PHI unless your organization has approved that use of Claude and this extension. Read the data-handling details.
These checks describe the supplied file. They cannot establish complete cohort coverage, consent, clinical appropriateness, or whether an outreach action happened.
Found a feed issue that blocks a launch?
A focused diagnostic connects the evidence to an owner, a decision, and a practical fix.