Retrieve
Answers are built from a finite, local corpus you can inspect, not from a model's memory.
Open source · MIT licensed · MCP-native
medground pairs a capable model with a citable corpus and a strict grounding check. Every answer cites a real source you can open.
MIT licensed. Python 3.11+. Runs locally.
Which targeted therapy for EGFR-mutant NSCLC?
Osimertinib improved progression-free survival in EGFR exon 19 deletion / L858R NSCLC.
LAURA trial · civic:eid12144
check_grounding passed · 1 / 1 claims cite a real source.
How it works
If a claim has no real source, it does not ship.
Answers are built from a finite, local corpus you can inspect, not from a model's memory.
Every claim carries a paper ID that links back to a real source.
A deterministic check confirms each citation exists before the answer reaches you.
The /doc skills
An illustrative case run through the medground MCP. Each /doc skill adds a grounded lens, and every finding here was checked by check_grounding. Hover a slice to follow the case.
Intake /doc-case
Builds a Patient Context Block: lung adenocarcinoma, EGFR exon 19 deletion, one prior line (osimertinib), ECOG 1. It steers every later query.
structured profile
Six of sixteen skills shown. Type /doc to route by intent, or /doc-help for the full menu. Illustrative case, decision support, not medical advice.
# clone and install
git clone https://github.com/ArnaudTurn-pro/medgroundAI
cd medgroundAI
uv sync
# build a corpus
uv run medground ingest civic
# connect it to Claude
claude mcp add medground -- uv run --directory . medground-mcp
# optional: install the /doc slash commands
./install-skills.sh
Then ask Claude a grounded question, or type /doc to pick a skill. No tool names required.
medground surfaces what the published literature says. It is not clinically validated and makes no clinical decisions. A grounded citation means a source is real and retrievable, not that it is correct for a patient. That judgment is always a clinician's.