heor-agent-mcp

HEORAgent MCP Server is an AI-driven health economics and outcomes research (HEOR) agent provided in the form of a Model Context Protocol (MCP) server, hosted on GitHub. It is aimed at pharmaceutical, biotechnology, CRO, and medical affairs teams, with the goal of automating and standardizing HEOR workflows.

Purpose The tool is used to support common tasks in health economics and outcomes research, including literature reviews across multiple data sources, risk of bias assessment, EQ-5D value set impact estimation, cost-effectiveness modeling, and preparation of submission materials for multiple HTA agencies. It emphasizes auditability and process rigor, making it suitable for scenarios that require the formation of HTA submission appendices.

Main capabilities

• Literature search: literature_search can search 44 data sources in parallel, and each call returns a source selection table showing which sources were used and why, facilitating HTA audit trails. Data sources cover PubMed, ClinicalTrials.gov, bioRxiv/medRxiv, ChEMBL, Wiley Online Library, WHO GHO, World Bank, OECD Health, IHME GBD, All of Us, FDA Orange Book/Purple Book, NICE TAs, CADTH, ICER, PBAC, G-BA AMNOG, IQWiG, HAS, AIFA, TLV, INESSS, CMS NADAC, PSSRU, NHS National Cost Collection, BNF, PBS Schedule, DATASUS, CONITEC, ANVISA, PAHO, IETS, FONASA, HITAP, ISPOR, OHE, EuroQol Group, etc., across 10 categories in total.
• Cost-effectiveness modeling: cost_effectiveness_model supports multi-state Markov models (default) or partitioned survival analysis (oncology), follows ISPOR good practices and NICE reference case (3.5% discount rate, half-cycle correction). It includes PSA (1,000–10,000 Monte Carlo iterations, probability of being cost-effective at WTP thresholds), OWSA (one-way sensitivity analysis and tornado diagram summary), CEAC (cost-effectiveness acceptability curve), EVPI (expected value of perfect information), as well as WTP assessment for NHS (£25–35K/QALY, updated April 2026), US payers ($100–150K), and social thresholds.
• Risk of bias assessment: automatically selects the appropriate Cochrane tool (RoB 2 / ROBINS-I / AMSTAR-2) based on study_type, returns a rob_results object that can be passed directly to hta_dossier_prep, replacing heuristic RoB estimates in GRADE tables with structured domain judgments.
• HTA submission material preparation: hta_dossier_prep can draft submission-ready sections for six HTA frameworks and includes gap analysis, and can accept the outputs of literature_search and cost_effectiveness_model.
• Pharmacovigilance and workflow orchestration: pv_classify can classify planned studies into EMA pharmacovigilance regulatory categories (PASS mandatory/voluntary, PAES, RMP Annex 4, DUS, active surveillance registry, pregnancy registry, spontaneous reporting, ICH E2E plan), returning the matching GVP module, ENCePP protocol template ID, RMP impact, FDA analogue, and submission obligations; it is a pure decision tree with a <200ms response. hta_dossier supports a pharmacovigilance plan section and can accept the classification results from pv_classify. maic_workflow can run the complete MAIC discovery + screening pipeline in a single call (ITC feasibility + parallel literature_search + screening + RoB + network).
• Living evidence intelligence (v1.17.0–v1.23.0): RWE and real-world safety (rwe.method_select study design selection, pv.comparative_safety class-level FAERS-style AE ranking, evidence.triangulation RCT↔RWE consistency), governed social listening (rwe.social_listening_protocol + pv.social_listening_triage, GVP Module VI ICSR triage, no crawling), single source of truth (evidence.claim_registry imports data in one go, evidence.consistency_check detects drift, publication.draft reuses declarations), living orchestration (evidence.gap_analysis, workflow.living_evidence, hta.living_gvd).
• AI transparency disclosure (v1.13.0, aligned with ISPOR ELEVATE-GenAI): 16 tools accept the ai_disclosure_level parameter; HTA/regulatory tools default to "submission", analytical tools default to "standard"; can be globally overridden via the environment variable HEORAGENT_DISCLOSURE_LEVEL=off|standard|submission.
• Output formats: all tools support output_format: text (default, Markdown tables and headings), json (structured object), docx (Microsoft Word document, saved to disk and returns the path). DOCX is saved to ~/.heor-agent/projects/{project}/reports/ (when a project is set) or ~/.heor-agent/reports/ (global).
• Audit records: each tool call returns a complete audit record, including a source selection table, queried sources, inclusion/exclusion counts and reasons, methodology (PRISMA-style for literature, ISPOR/NICE for economics), assumptions and rationale, and warnings. Suitable for inclusion in HTA submission appendices.
• Knowledge base: projects are located at ~/.heor-agent/projects/{project-id}/, containing raw/literature/, raw/models/, raw/dossiers/, reports/, wiki/ (manually curated Obsidian-compatible markdown, supporting [[wikilinks]]). Pass project: "project-id" to any tool to automatically save results.

Usage

• Client integration: supports Claude Code, Claude Desktop / claude.ai Desktop, Cursor, Continue, Cline, all using stdio. Configuration: for Claude Code, add and restart; for Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) and restart; for Cursor, go to Settings → MCP → Add new MCP server; for Continue, under the mcpServers key in ~/.continue/config.json; for Cline, go to Settings → MCP Servers → Edit MCP Settings.
• Quick verification: open two terminal tabs, start the server in HTTP mode in Tab 1, and confirm the response in Tab 2. If you see the expected JSON output, the npm package is working; if you get command not found, run node --version to confirm Node ≥20; for other errors, you can https://github.com/neptun2000/heor-agent-mcp/issues submit an issue. Actual integration uses stdio, so HTTP mode is not needed.
• Hosted usage: HEORAgent is available on ChatGPT (Plus / Team); type /heor to use it, applicable to any conversation; the Web UI (Claude, BYOK) is located at https://web-michael-ns-projects.vercel.app，需自备 Anthropic API key, running the full v1.6.3 toolset.
• Custom GPT: v1.0.4 provides an OpenAPI 3.1 adapter /api/openapi, allowing quick building of a Custom GPT.
• Installation command: npx heor-agent-mcp (for Claude Desktop / Claude Code).

Use cases Suitable for pharmaceutical, biotechnology, CRO, and medical affairs teams, for literature reviews, indirect comparisons (Bucher), MAIC population-adjusted comparisons, cost-effectiveness analysis, budget impact analysis, survival curve fitting, evidence networks and NMA feasibility assessment, risk of bias assessment, EQ-5D 5L impact estimation, and preparation of submission materials for EU JCA and institutions such as NICE/ICER. The documentation provides multiple copy-paste example prompts covering single-tool and multi-tool workflows.

Limitations The materials do not provide specific fee and pricing information, so it is not available for now. Using the Web UI requires your own Anthropic API key. ChatGPT integration requires a Plus / Team / Enterprise account. Node version must be ≥20. Some data sources may require API keys (the documentation indicates warning markers for missing API keys).