Gemini Setup Guide

For students and teams using Google Gemini instead of Claude

When to Use This Guide

Use this guide if your organization has Google Gemini licenses but not Claude access. All MCP servers in this repo work with both Claude and Gemini — the servers don't care which AI model calls them.

Three setup paths (pick one):

Path	Best For	Requirements
A. VSCode + Gemini Code Assist	Corporate laptops with Gemini licenses	VSCode, Gemini Code Assist extension
B. Gemini CLI (terminal)	Developers who prefer the terminal	Node.js 18+, `@google/gemini-cli`
C. Streamlit Student App	Browser-based, no local MCP config	GCP project for Cloud Run deployment

Prerequisites (All Paths)

# 1. Python 3.11+ and git
python3 --version   # must be 3.11+
git --version

# 2. Install uv (Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 3. Clone the repo
git clone https://github.com/lynnlangit/precision-medicine-mcp.git
cd precision-medicine-mcp

# 4. Install at least one MCP server to test with
cd servers/mcp-mockepic && uv sync && cd ../..

Path A: VSCode + Gemini Code Assist

This is the most common setup for corporate environments with existing Gemini licenses.

Step 1: Install Extensions

In VSCode, install:

Google Cloud Code extension
Gemini Code Assist extension

Step 2: Configure MCP Servers

Create or edit ~/.gemini/settings.json:

{
  "mcpServers": {
    "mockepic": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/ABSOLUTE/PATH/TO/precision-medicine-mcp/servers/mcp-mockepic",
        "python",
        "-m",
        "mcp_mockepic"
      ],
      "env": {
        "EPIC_DRY_RUN": "true"
      }
    }
  }
}

Replace /ABSOLUTE/PATH/TO/ with your actual path (e.g., /Users/yourname/Documents/GitHub/).

Corporate Mac tip: Use absolute paths for the uv command too if it's not on your system PATH. Find it with which uv and use the full path (e.g., /Users/yourname/.local/bin/uv).

Step 3: Enable Agent Mode

This is the critical step. Without Agent Mode, Gemini will describe tools instead of calling them.

Open the Gemini side panel in VSCode
Look for the Agent or Spark toggle — it must be active (highlighted)
If you don't see the toggle, check VSCode Settings > search for gemini > look for agent-related settings

Note: The exact setting name may vary by extension version. Check the Gemini Code Assist extension settings UI rather than editing JSON directly.

Step 4: Verify

Type /mcp in the Gemini chat panel — your server should show CONNECTED
Test with an explicit prompt:

Use the mockepic tool to get patient demographics for patient-001. Execute now.

Expected: Gemini calls the MCP tool and returns Sarah Anderson's demographics (58yo, Stage IV HGSOC).

If Gemini describes the tool instead of calling it: You're still in Chat Mode, not Agent Mode. Go back to Step 3.

Adding More Servers

Once mockepic works, add servers to ~/.gemini/settings.json one at a time. Start with these for the PatientOne workflow:

{
  "mcpServers": {
    "mockepic": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/precision-medicine-mcp/servers/mcp-mockepic", "python", "-m", "mcp_mockepic"],
      "env": { "EPIC_DRY_RUN": "true" }
    },
    "spatialtools": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/precision-medicine-mcp/servers/mcp-spatialtools", "python", "-m", "mcp_spatialtools"],
      "env": { "SPATIAL_DATA_DIR": "/ABSOLUTE/PATH/TO/precision-medicine-mcp/data", "SPATIAL_DRY_RUN": "true" }
    },
    "multiomics": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/precision-medicine-mcp/servers/mcp-multiomics", "python", "-m", "mcp_multiomics"],
      "env": { "MULTIOMICS_DATA_DIR": "/ABSOLUTE/PATH/TO/precision-medicine-mcp/data/multiomics", "MULTIOMICS_DRY_RUN": "true" }
    },
    "fgbio": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/precision-medicine-mcp/servers/mcp-fgbio", "python", "-m", "mcp_fgbio"],
      "env": { "FGBIO_DRY_RUN": "true" }
    }
  }
}

For the full list of all servers and their environment variables, see the Claude Desktop config template — the server names, args, and env vars are identical for Gemini.

Path B: Gemini CLI (Terminal)

The Gemini CLI is a terminal-based alternative, similar to Claude Code. FastMCP has built-in support for installing servers into it.

Step 1: Install Gemini CLI

npm install -g @google/gemini-cli
gemini --version   # verify installation

Step 2: Install MCP Servers

FastMCP can register servers directly:

cd servers/mcp-mockepic
uv run fastmcp install gemini-cli src/mcp_mockepic/server.py \
  --name mockepic \
  --env EPIC_DRY_RUN=true

Or manually edit ~/.gemini/settings.json using the same JSON format from Path A above.

Step 3: Verify

gemini   # launches the Gemini CLI
# Then type: "Use mockepic to get demographics for patient-001"

Path C: Streamlit Student App (Browser)

The repo includes a pre-built student web app that uses Gemini and connects to MCP servers deployed on GCP Cloud Run. No local MCP configuration needed — students just open a browser.

Requirements: A GCP project with MCP servers deployed to Cloud Run (see the deployment guide).

Setup

cd ui/streamlit-app-students
cp .env.example .env
# Edit .env: add your GEMINI_API_KEY
# Set USE_MOCK_MCP=true for weeks 1-2

pip install -r requirements.txt
streamlit run app.py

Safety guardrails (built in):

Max 4,096 tokens per request
Max 50,000 tokens per session (~$1.50)
Max 50 requests per session

See the student app .env.example for all configuration options.

Corporate Mac Troubleshooting

Server won't connect

Symptom: /mcp shows server not connected, or no servers listed.

Fixes (try in order):

Absolute paths: Replace uv with output of which uv in your settings.json
Restart VSCode: MCP config changes require a restart
Check Python: Run uv run --directory /path/to/server python -m mcp_mockepic in terminal first — if it fails there, fix that before debugging VSCode

Tools listed but not executed

Symptom: Gemini says "I can see the mockepic tool has these functions..." but doesn't call them.

Fixes:

Agent Mode: Toggle must be active (Step 3 above)
Explicit prompts: Use action-oriented language:
- Instead of: "What data is available for PatientOne?"
- Try: "Use the mockepic tool to get patient demographics for patient-001. Execute now."
Model selection: Use Gemini 2.5 Flash or newer — older models have weaker tool-calling

Proxy blocks MCP server

Symptom: Server starts but can't reach external APIs, or connection times out.

Fix: MCP servers run as child processes of VSCode. If your corporate proxy requires configuration, set proxy env vars in your settings.json:

{
  "mcpServers": {
    "mockepic": {
      "command": "uv",
      "args": ["..."],
      "env": {
        "EPIC_DRY_RUN": "true",
        "HTTPS_PROXY": "http://your-corporate-proxy:8080"
      }
    }
  }
}

Complex JSON output causes errors

Symptom: Tool fires but Gemini shows "improper format" or gets stuck in a loop.

Fix: Set DRY_RUN=true for initial testing — DRY_RUN responses are simpler and less likely to trigger Gemini's JSON parsing issues. Once basic connectivity works, switch to DRY_RUN=false for real analysis.

Study Group Quick Checklist

For instructors setting up a study group on corporate Macs:

Before session 1: Create a standardized settings.json with absolute paths for your team's machines
Session 1 (30 min): Everyone installs extensions, copies config, verifies /mcp shows CONNECTED
Session 1 test: Everyone runs "Use mockepic to get demographics for patient-001. Execute now." and gets a result
Weeks 1-2: Use DRY_RUN=true (free practice, synthetic data)
Week 3+: Switch to DRY_RUN=false for real analysis

Gemini vs Claude: What's Different

Feature	Claude (Code/Desktop)	Gemini (VSCode/CLI)
MCP tool calling	Aggressive — calls tools automatically	Requires Agent Mode + explicit prompts
Config file	`claude_desktop_config.json`	`~/.gemini/settings.json`
Server format	Identical `uv run --directory` syntax	Identical `uv run --directory` syntax
Env vars	Identical per-server DRY_RUN vars	Identical per-server DRY_RUN vars
Best model for tools	Claude Sonnet 4.5 / Opus 4.6	Gemini 2.5 Flash / 2.5 Pro
Streamlit app	Uses Claude provider	Uses Gemini provider (built in)

The MCP servers are identical — only the AI client configuration differs.

Navigation Menu

Gemini Setup Guide

When to Use This Guide

Prerequisites (All Paths)

Path A: VSCode + Gemini Code Assist

Step 1: Install Extensions

Step 2: Configure MCP Servers

Step 3: Enable Agent Mode

Step 4: Verify

Adding More Servers

Path B: Gemini CLI (Terminal)

Step 1: Install Gemini CLI

Step 2: Install MCP Servers

Step 3: Verify

Path C: Streamlit Student App (Browser)

Setup

Corporate Mac Troubleshooting

Server won't connect

Tools listed but not executed

Proxy blocks MCP server

Complex JSON output causes errors

Study Group Quick Checklist

Gemini vs Claude: What's Different

Related Documentation

FilesExpand file tree

gemini-setup.md

Latest commit

History

gemini-setup.md

File metadata and controls

Gemini Setup Guide

When to Use This Guide

Prerequisites (All Paths)

Path A: VSCode + Gemini Code Assist

Step 1: Install Extensions

Step 2: Configure MCP Servers

Step 3: Enable Agent Mode

Step 4: Verify

Adding More Servers

Path B: Gemini CLI (Terminal)

Step 1: Install Gemini CLI

Step 2: Install MCP Servers

Step 3: Verify

Path C: Streamlit Student App (Browser)

Setup

Corporate Mac Troubleshooting

Server won't connect

Tools listed but not executed

Proxy blocks MCP server

Complex JSON output causes errors

Study Group Quick Checklist

Gemini vs Claude: What's Different

Related Documentation