Heor Agent MCP

What neptun2000/heor-agent-mcp MCP server does

The neptun2000/heor-agent-mcp MCP server exposes an MCP toolset for health economics and outcomes research. Its capabilities cover the workflow from finding studies and screening abstracts through bias assessment, evidence-network construction, comparative effectiveness analysis, economic modeling, and HTA submission drafting.

Literature searches can query 44 medical data sources and produce a PRISMA-style audit trail. Evidence tools support PICO relevance scoring, study-design classification, Cochrane RoB 2, ROBINS-I, AMSTAR-2, GRADE-related summaries, Bucher comparisons, frequentist network meta-analysis, MAIC, STC, and feasibility checks for indirect treatment comparisons.

Economic analysis includes parametric survival fitting, Markov models, partitioned survival models, decision trees, probabilistic and one-way sensitivity analysis, cost-effectiveness acceptability curves, EVPI, EVPPI, QALYs, and evLYG. Other tools cover ISPOR-oriented budget impact modeling, EQ-5D value-set references, and drafts for NICE, EMA, FDA, IQWiG, HAS, and EU JCA submissions.

How it works

The server runs as a local Node.js MCP process and can communicate with MCP hosts over standard input/output. The README also documents an HTTP mode for local health checks and testing. MCP clients call individual tools with structured inputs; some workflows can be chained, such as searching literature, screening results, assessing risk of bias, building an evidence network, and passing outputs into an HTA dossier.

Project tools provide a persistent workspace. A project can be initialized, files in its raw or wiki trees can be searched and read, and compiled evidence can be written to an Obsidian-compatible wiki. Citation URLs can be checked before results are presented. The README describes a current release with additional workflow, pharmacovigilance, RWE, disclosure, and publication-related tools beyond the core list supplied here.

Setup and configuration

The neptun2000/heor-agent-mcp MCP server requires Node.js 20 or later according to the installation notes. Install or run it with npx heor-agent-mcp. Claude Desktop uses an MCP configuration entry with npx as the command and heor-agent-mcp as its argument. The README also gives client-specific setup guidance for Claude Code, Cursor, Continue, and Cline.

For a local HTTP test, the documented command sets MCP_HTTP_PORT=8080 and starts the package. A request to /health should return a status response identifying the server and version. The HTTP test mode is separate from the stdio configuration used for the normal Claude and other MCP-client integrations.

The optional HEORAGENT_DISCLOSURE_LEVEL setting can override the default AI disclosure mode with off, standard, or submission. The hosted web UI is a separate access path and requires the user's Anthropic API key; the README identifies literature search and link validation as usable without API keys.

Tools and capabilities

• Search medical literature with source-specific retrieval and audit information.
• Screen abstracts using PICO relevance and study-design classification.
• Assess risk of bias with RoB 2, ROBINS-I, or AMSTAR-2 frameworks.
• Evaluate and perform Bucher, NMA, MAIC, and STC comparisons.
• Fit survival curves and run cost-effectiveness or budget-impact models.
• Prepare HTA evidence and dossier drafts for named regulatory and payer contexts.
• Maintain searchable project evidence in raw and wiki file trees.

The tool inputs and outputs should be reviewed by qualified HEOR, clinical, statistical, and regulatory specialists before being used in a submission or decision. Model results depend on the assumptions and evidence supplied to each call.