helix-mcp 1.5.0

Clinical evidence synthesis MCP server — ClinicalTrials.gov, PubMed, and openFDA in one call

🧬 Helix

Clinical evidence synthesis engine — free, no API key, production-grade.

Give any AI model — or any HTTP client — structured, scored access to the world's three largest free health databases. In a single call.

400,000+ Clinical Trials	35M+ PubMed Papers	FDA Drug Labels	< 4s Response
ClinicalTrials.gov live	Full abstracts via efetch	openFDA drug information	All 3 databases in parallel

---

See It

curl -X POST http://localhost:8000/synthesize \
  -H "Content-Type: application/json" \
  -d '{"condition": "T2D", "age": 45, "sex": "MALE"}'

{
  "clinicalInsight": {
    "condition": "Type 2 Diabetes",
    "expanded_from": "T2D",
    "total_trials": 16,
    "top_score": 94.0,
    "average_score": 77.66
  },
  "trialProfiles": [
    {
      "id": "NCT05099770",
      "title": "PROACT: A Study of REACT in Subjects With Type 2 Diabetes",
      "phase": ["PHASE3"],
      "final_score": 94.0,
      "score_vector": {
        "condition_match": 1.0,
        "eligibility_fit": 0.8,
        "evidence_support": 1.0,
        "trial_phase_maturity": 1.0
      },
      "risk_flags": []
    }
  ],
  "excludedTrials": [
    {
      "id": "NCT00000042",
      "title": "Pediatric Glucose Management Study",
      "exclusion_reason": "age 45 above max 18"
    }
  ]
}

Every trial gets a score vector showing exactly why it ranked where it did. Ineligible trials appear in excludedTrials with a precise rejection reason — they never silently disappear.

---

Install

pip install helix-mcp

REST API

helix-api
# → http://localhost:8000/docs

MCP Server (Claude Desktop)

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "helix": { "command": "helix" }
  }
}

Then ask Claude: "Find clinical trials for a 52-year-old female with NSCLC in London"

Docker

docker compose up

---

How It Works

When you call synthesize_evidence, Helix:

1. Expands medical abbreviations — T2D → Type 2 Diabetes, NSCLC → Non-Small Cell Lung Cancer (70+ mappings)
2. Resolves conditions to authoritative NLM MeSH terms before querying PubMed — the same vocabulary PubMed uses internally
3. Queries ClinicalTrials.gov, PubMed, and openFDA concurrently
4. Scores every trial using a BM25 relevance index built across the full trial corpus
5. Returns ranked profiles with a four-component score_vector and an explainability_vector showing the raw numbers behind each score

Condition input
      ↓
synonym expansion → MeSH resolution
      ↓
ClinicalTrials.gov ──┐
PubMed (MeSH query) ─┼── parallel asyncio.gather
openFDA ─────────────┘
      ↓
BM25 corpus scoring across all trials
      ↓
hard eligibility gate (age + sex)
      ↓
ranked profiles + excluded trials + clinical insight

---

Scoring Formula

final_score = 100 × (
    0.35 × condition_match        # BM25 relevance: condition vs trial corpus
  + 0.30 × eligibility_fit        # age-window centrality (1.0=center, 0.5=edge)
  + 0.20 × evidence_support       # fraction of PubMed papers supporting condition
  + 0.15 × trial_phase_maturity   # Phase 3/4=1.0, Phase 2=0.6, Phase 1=0.3
)

---

Tools

Tool	Description
synthesize_evidence	Full cross-database synthesis — scored, ranked, explained
find_trials	Search ClinicalTrials.gov with condition, location, sex, phase
search_papers	Search PubMed with MeSH-resolved queries, full abstracts
lookup_drug	FDA drug information by brand or generic name
match_eligibility	Match a patient profile to trials ranked by eligibility fit
health_check	Live latency check against all three upstream APIs

All tools accept medical abbreviations. All tools are cached. All tools never raise.

---

Data Sources

Source	Access
ClinicalTrials.gov	Free, no key
PubMed E-utilities	Free, email optional
openFDA	Free, no key
NLM MeSH API	Free, no key

---

Development

git clone https://github.com/Al1Abdullah/Helix.git
cd Helix
pip install -e ".[dev]"
pytest

---

License

MIT — see LICENSE