75% of US adults take dietary supplements. Most assume "natural" means safe. AI health agents are reinforcing that assumption because they have no interaction data in the loop. Here's how to fix that.
There's a growing wave of AI health agents. Chatbots, supplement advisors, wellness copilots, nutrition coaches. They're built on powerful language models that can synthesize research, explain mechanisms, and give personalized-sounding advice.
But they're all missing the same thing: a real-time interaction safety layer.
The Problem Is Infrastructure, Not Intelligence.
When someone asks an AI "should I take magnesium?" the model draws on training data. It might know that magnesium supports muscle function, sleep, and cardiovascular health. It probably gives a reasonable answer.
But when that same person is on lisinopril, or metformin, or warfarin, the answer changes completely. And the model doesn't check. It can't. There's no database in the loop.
This isn't a hallucination problem. The model isn't making things up. It's an infrastructure problem. The model has no access to structured interaction data at inference time.
What "Real" Interaction Data Looks Like
A useful interaction database isn't just a list of "don't mix X with Y." It needs:
Directionality. Magnesium affecting warfarin absorption is different from warfarin affecting magnesium levels. Interactions are directed, not symmetric.
Severity scoring. Not all interactions are equal. St. John's Wort with SSRIs is potentially life-threatening (serotonin syndrome). Calcium with iron is inconvenient (reduced absorption). Your agent needs to know the difference.
Evidence quality. Is this interaction backed by randomized controlled trials, case reports, or theoretical mechanisms? The confidence level matters for how strongly the agent should flag it.
Adverse event signals. The FDA's FAERS and CAERS databases contain real-world adverse event data. If 47 people reported liver problems after taking a specific supplement, that's a signal worth surfacing, even before formal studies exist.
Mechanism of action. "Reduces absorption" vs "induces CYP3A4 enzyme" vs "competes for same receptor." The mechanism determines whether timing can mitigate the interaction or whether it's a hard contraindication.
How We Built TruthStack
TruthStack is a supplement-drug interaction API built for exactly this use case. The database currently contains:
1,008 directed interactions with severity and evidence scoring
805 FDA FAERS adverse event signals
426 CAERS pharmacovigilance signals with PRR (Proportional Reporting Ratio) analysis
235 research findings sourced from PubMed and ClinicalTrials.gov
31 compound safety profiles with organ-system risk flags
25 drug profiles covering the most common medications
95 compounds with 584 aliases for fuzzy matching
3,510 total data points
The data comes from three automated pipelines:
Research agent — scans PubMed and ClinicalTrials.gov daily for new supplement research, extracts interaction findings, and runs them through an auto-review pipeline before deployment.
FAERS agent — pulls adverse event reports from the FDA's drug-focused reporting system, identifies supplement-drug co-occurrence patterns, and flags statistically significant signals.
CAERS agent — queries the FDA's supplement-specific adverse event database (CFSAN Adverse Event Reporting System), computes PRR for each reaction, and generates compound safety profiles with organ-system alerts (hepatic, cardiac, bleeding, renal).
Every interaction in the database has a source, an evidence grade, and a severity score. No black boxes.
Two Ways to Integrate
REST API
Standard HTTPS endpoints. Works with any language, any framework.
bashcurl -X POST https://api.truthstack.co/api/interactions/check \
-H "Content-Type: application/json" \
-H "X-API-Key: your-key" \
-d '{
"supplements": ["ashwagandha", "fish oil", "magnesium"],
"medications": ["sertraline"]
}'
The response includes every interaction found, severity scores, evidence quality, mechanisms, and an overall risk level for the stack.
Full docs: api.truthstack.co/docs
MCP Server
If you're building with Claude Desktop, LangChain, or any MCP-compatible framework, the TruthStack MCP server gives your agent 6 structured tools:
check_interactions — check a full supplement + medication stack
search_compounds — fuzzy search across 584 aliases
get_compound_info — detailed compound profiles
explain_interaction — mechanism and clinical context
get_evidence — FAERS data and research citations
get_safety_signals — CAERS adverse event signals with PRR analysis
Install via npm:
bashnpm install -g truthstack-mcp
Or add to Claude Desktop config:
json{
"mcpServers": {
"truthstack": {
"command": "npx",
"args": ["truthstack-mcp"],
"env": {
"VAULT_API_URL": "https://api.truthstack.co",
"VAULT_API_KEY": "your-key"
}
}
}
}
GitHub: github.com/TruthStack1/truthstack-mcp
What the Data Actually Catches
Some real examples from the database:
Red yeast rice + statins: Red yeast rice contains monacolin K, which is chemically identical to lovastatin. Taking it alongside a prescribed statin doubles the effective statin dose. The CAERS data shows a strong myalgia signal (PRR = 25.37). Most AI agents would miss this because they don't know red yeast rice IS a statin.
Green tea extract + liver risk: The CAERS data shows moderate signals for elevated ALT, bilirubin, and jaundice. This is a known hepatotoxicity risk with concentrated extracts (not regular tea). An agent recommending green tea extract for weight loss should be flagging this.
Ashwagandha + SSRIs: Theoretical serotonergic interaction. Not life-threatening like St. John's Wort, but worth flagging, especially at high doses. The agent needs the severity context to calibrate its response.
Fish oil + blood thinners: Additive anticoagulant effect. The interaction is real but the clinical significance depends on dose. A good interaction database provides the mechanism so the agent can reason about it rather than just saying "avoid."
Why This Matters Now
The supplement market is $60B+ and growing. 75% of US adults take at least one supplement. Most people assume "natural" means safe, and most AI agents reinforce that assumption because they have no data to contradict it.
Meanwhile, AG1's clinical trial just got publicly challenged by Bryan Johnson, highlighting exactly the gap between marketing claims and evidence-based safety data.
If you're building anything that touches supplement recommendations, you need an interaction safety layer. Not because regulators require it (yet), but because your users are making real health decisions based on your agent's output.
Get Started
API docs: api.truthstack.co/docs
MCP server: github.com/TruthStack1/truthstack-mcp
LangChain wrapper: github.com/TruthStack1/truthstack-langchain
Example agent: github.com/TruthStack1/truthstack-example-health-agent
We're looking for the first 10 health AI builders to integrate. If your agent recommends supplements, DM me or open an issue on GitHub.
Top comments (0)