Biomedical Agent

Name: biomed-agent
Availability: InStock
Rating: 3.0 (10 reviews)

General-purpose biomedical knowledge assistant that connects to multiple MCP biomedical data sources and synthesizes answers with confidence, citations, and limitations.

What It Does

Connects to biomedical MCP servers (OpenTargets, Monarch, MyGene, MyChem, MyDisease)
Routes questions across tools and returns synthesized answers
Returns structured metadata for trust:
- confidence (high|medium|low)
- citations (observation ID + tool)
- limitations (missing evidence, conflicts, uncertainty)
Supports both CLI and Streamlit interfaces

Architecture

%%{init: {'flowchart': {'curve': 'linear', 'nodeSpacing': 45, 'rankSpacing': 60}}}%%
flowchart LR
    classDef ui fill:#E8F0FE,stroke:#1A73E8,color:#0B3D91,stroke-width:1.5px;
    classDef core fill:#E6F4EA,stroke:#188038,color:#0D652D,stroke-width:1.5px;
    classDef data fill:#E0F7FA,stroke:#00838F,color:#004D40,stroke-width:1.5px;
    classDef output fill:#FFF4E5,stroke:#F9AB00,color:#8D5B00,stroke-width:1.5px;
    classDef neutral fill:#F8F9FA,stroke:#5F6368,color:#202124,stroke-width:1.2px;

    subgraph UX[Interfaces]
        direction TB
        U["User"]
        CLI["CLI<br/>ui/cli.py"]
        WEB["Streamlit UI<br/>ui/app.py"]
    end

    subgraph CORE[Orchestration]
        direction TB
        AGENT["BiomedAgent<br/>core/agent.py"]
        GUARD["Guardrails<br/>risk + safety checks"]
        LOOP["Reasoning loop<br/>LLM -> JSON action"]
        REG["Tool registry"]
        SYN["Synthesis + metadata<br/>confidence/citations/limitations"]
    end

    subgraph MCPLAYER[MCP Layer]
        direction TB
        MCP["MCPClient<br/>core/mcp_client.py"]
        OT["opentargets-mcp"]
        MO["monarch-mcp"]
        MG["mygene-mcp"]
        MC["mychem-mcp"]
        MD["mydisease-mcp"]
    end

    subgraph OUT[Outputs]
        direction TB
        RESP["Answer"]
        META["confidence + citations + limitations"]
    end

    U --> CLI
    U --> WEB
    CLI --> AGENT
    WEB --> AGENT

    AGENT --> GUARD --> LOOP
    LOOP --> REG --> MCP
    MCP <--> OT
    MCP <--> MO
    MCP <--> MG
    MCP <--> MC
    MCP <--> MD
    MCP -- tool results --> LOOP
    LOOP --> SYN

    SYN --> RESP
    SYN --> META

    class U neutral
    class CLI,WEB ui
    class AGENT,GUARD,LOOP,REG,SYN core
    class MCP,OT,MO,MG,MC,MD data
    class RESP,META output

Query Lifecycle

%%{init: {'flowchart': {'curve': 'linear', 'nodeSpacing': 35, 'rankSpacing': 45}}}%%
flowchart LR
    classDef input fill:#E8F0FE,stroke:#1A73E8,color:#0B3D91,stroke-width:1.2px;
    classDef retrieve fill:#E6F4EA,stroke:#188038,color:#0D652D,stroke-width:1.2px;
    classDef output fill:#FFF4E5,stroke:#F9AB00,color:#8D5B00,stroke-width:1.2px;

    Q["1. User question"] --> R["2. Risk/safety check"]
    R --> T["3. Select tools"]
    T --> C["4. Execute MCP calls"]
    C --> O["5. Normalize observations"]
    O --> A["6. Compose answer"]
    A --> M["7. Add metadata<br/>confidence + citations + limitations"]
    M --> X["8. Return response to CLI/UI"]

    class Q,R input
    class T,C,O,A retrieve
    class M,X output

Safety Scope

Intended for research and educational use.
Not a clinical decision system.
Does not provide personalized diagnosis or treatment advice.

Prerequisites

Python 3.12 or 3.13
uv for dependency management
MCP servers installed locally:

Setup

Install dependencies:

uv sync

Configure environment variables:

cp .env.example .env

Edit .env with your API key and local MCP server paths.

CLI Usage

List server availability:

uv run python -m ui.cli list-servers

List tools:

uv run python -m ui.cli list-tools
uv run python -m ui.cli list-tools --server opentargets --server mychem
uv run python -m ui.cli list-tools --capability drug

Run a biomedical query:

uv run python -m ui.cli query "What is the mechanism of action of vemurafenib?"

Run a query and show trace:

uv run python -m ui.cli query "Which genes are associated with Parkinson disease?" --show-steps

Run a query and output full JSON:

uv run python -m ui.cli query "What drugs target BRAF?" --json

Call one tool directly:

uv run python -m ui.cli call-tool opentargets.search_entities '{"query_string": "BRAF", "entity_names": ["target"]}'

Interactive chat:

uv run python -m ui.cli chat

Streamlit Usage

uv run streamlit run ui/app.py

Then:

Select available servers in the sidebar.
Click Connect.
Ask questions in Chat or Direct Query.
Inspect Sources and Limitations in each response.

biomed-agent