Knowledge Graph Causal Discovery MCP Server

> View on ApifyForge | Use on Apify Store

---

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf):

{
  "mcpServers": {
    "knowledge-graph-causal-discovery-mcp": {
      "url": "https://ryanclinton--knowledge-graph-causal-discovery-mcp.apify.actor/mcp"
    }
  }
}

---

Knowledge graph causal discovery over multi-domain research data, delivered through a single Model Context Protocol interface. This MCP server is built for researchers, data scientists, and AI agents that need to go beyond correlation — discovering directed causal structure, estimating treatment effects, and reasoning about counterfactuals from the published literature and public datasets.

The server orchestrates 17 Apify actors in parallel across five source domains — academic, biomedical, regulatory, economic, and safety — assembling the results into a unified causal knowledge graph. Eight specialized tools then apply rigorous causal inference algorithms: FCI skeleton learning, GES with BIC scoring, Pearl's do-calculus with the ID algorithm, twin network counterfactuals, TMLE estimation, RotatE knowledge graph embeddings, sheaf cohomology consistency checking, and Shapley source attribution. Every tool call returns structured JSON with mathematical scores and supporting evidence.

⬇️ What data can you access?

| Data Point | Source | Coverage |
|---|---|---|
| 📄 Academic papers and citations | OpenAlex, Semantic Scholar, Crossref | 250M+ scholarly works with citation graphs |
| 📑 Preprints and open access | arXiv, CORE | Physics, CS, quantitative biology, math |
| 🧬 Biomedical literature | PubMed | 36M+ citations with MeSH indexing |
| 🏥 Clinical trials | ClinicalTrials.gov | 450K+ registered studies with protocol data |
| 💊 Drug adverse event reports | OpenFDA | FDA FAERS pharmacovigilance database |
| 🔬 NIH research grants | NIH Reporter | Active and historical funded projects |
| 📜 Federal regulations | Federal Register | US regulatory actions and proposed rules |
| 🏛️ Congressional legislation | Congress.gov | Bills, resolutions, and amendments |
| 🗂️ Government datasets | Data.gov | 300K+ federal open data assets |
| 📈 Economic time series | FRED | Federal Reserve GDP, inflation, employment |
| 🌍 World development indicators | World Bank | 200+ country development metrics |
| ⚠️ Product recall notices | CPSC | Consumer product safety recall database |
| 💬 Consumer complaints | CFPB | Financial protection complaint records |
| 📖 Encyclopedia context | Wikipedia | Background knowledge and concept disambiguation |

Why use Knowledge Graph Causal Discovery MCP?

Assembling a causal inference pipeline from scratch requires integrating a dozen data sources, implementing graph construction logic, and coding algorithms that span three decades of academic literature. A typical research team spending a week on this still ends up with a pipeline that covers two or three data domains at best.

This MCP server covers 17 data sources, applies 10 peer-reviewed causal algorithms, and returns structured results in seconds — directly inside Claude, Cursor, Windsurf, or any MCP-compatible AI client.

- Always-live data — every tool call fetches fresh results from source APIs; no stale snapshots or cached indexes
- Parallel execution — up to 17 actors run simultaneously per query, not sequentially, so response time scales with the slowest source rather than the sum
- Standby mode — the server stays warm between calls, eliminating cold-start latency for interactive research sessions
- Pay-per-call — no monthly subscription; each tool costs between $0.035 and $0.050, so a full 8-tool pipeline costs under $0.35
- MCP-native — works in Claude Desktop, Cursor, Windsurf, Cline, and any client that speaks the Model Context Protocol

⬆️ MCP tools

| Tool | Price | Algorithm | Best for |
|------|-------|-----------|----------|
| discover_causal_structure | $0.045 | FCI + GES + additive noise model | Initial causal graph structure from observational data |
| compute_interventional_effects | $0.050 | Pearl's do-calculus + ID algorithm + Balke-Pearl LP | Policy evaluation, treatment planning, intervention design |
| simulate_counterfactuals | $0.045 | Twin network method + Tian-Pearl bounds | "What if" analysis, legal causation, necessity/sufficiency |
| extract_causal_claims_literature | $0.035 | NLP pattern matching + evidence classification | Systematic reviews, evidence synthesis, claim auditing |
| embed_causal_knowledge_graph | $0.040 | RotatE complex-valued embeddings | Link prediction, entity similarity, pathway discovery |
| estimate_causal_effect_tmle | $0.050 | TMLE + Super Learner ensemble + influence function CI | Semiparametric ATE estimation with doubly-robust CI |
| check_graph_consistency | $0.035 | Sheaf cohomology H¹(G,F) | Validating causal assumptions, identifiability checks |
| attribute_source_contribution | $0.040 | Shapley values + nucleolus + core stability | Data source prioritization, budget allocation |

Use cases for knowledge graph causal discovery

Drug safety signal detection

Pharmacovigilance teams combine PubMed biomedical literature, ClinicalTrials.gov outcome data, and FDA adverse event reports into a single causal graph. The discover_causal_structure tool identifies directed edges between compounds and adverse outcomes. The compute_interventional_effects tool estimates P(adverse event | do(prescribe drug)) using back-door adjustment on confounders sourced from NIH grant data and OpenAlex citations.

Policy impact assessment

Policy analysts estimate causal effects of regulatory interventions on economic outcomes by combining Federal Register rules, FRED economic time series, and World Bank development indicators. The estimate_causal_effect_tmle tool applies TMLE with Super Learner to produce doubly-robust average treatment effect estimates with 95% confidence intervals from the influence function — going beyond naive before/after comparison.

Systematic review and evidence synthesis

Literature reviewers use extract_causal_claims_literature to scan thousands of academic papers across OpenAlex, Semantic Scholar, Crossref, arXiv, and CORE simultaneously. Claims are classified by strength (strong/moderate/weak/correlational) and evidence level (RCT/observational/case study/review). Conflicting claims across sources are flagged automatically, replacing weeks of manual screening.

Counterfactual reasoning for legal and regulatory causation

Legal teams and regulators assessing causation in product liability or pharmaceutical harm cases use simulate_counterfactuals to compute the Probability of Necessity (PN = P(Y_x'=0 | X=x, Y=y)) and Probability of Sufficiency (PS = P(Y_x=1 | X=x', Y=0)) via the twin network method. Tian-Pearl monotonicity bounds are validated to constrain the counterfactual probabilities.

Knowledge graph completion in biomedical AI

AI research teams use embed_causal_knowledge_graph to generate RotatE complex-valued entity embeddings where relations are unit-modulus rotations in complex space (t = h · r, |r_i| = 1). MRR and Hits@10 link prediction metrics identify missing drug-disease or gene-pathway edges. Self-adversarial negative sampling with margin gamma ensures high-quality embeddings even in sparse graph regions.

Data acquisition prioritization

Research operations teams with limited budgets use attribute_source_contribution to calculate Shapley values for each data domain (academic, biomedical, regulatory, economic, safety). The Shapley allocation phi_i quantifies each source's marginal contribution to causal graph quality across all subsets. Nucleolus computation and core non-emptiness check confirm allocation stability before committing to data subscriptions.

How to connect this MCP server

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "knowledge-graph-causal-discovery": {
      "url": "https://knowledge-graph-causal-discovery-mcp.apify.actor/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_APIFY_TOKEN"
      }
    }
  }
}

Cursor / Windsurf / Cline

Add the MCP endpoint in your editor's MCP settings panel:

- Endpoint URL: https://knowledge-graph-causal-discovery-mcp.apify.actor/mcp
- Authentication: Bearer token with your Apify API token

Python (MCP client)

```python
import anthropic

client = anthropic.Anthropic()