Skip to content

GEMINI_SUPPORT.md

Latest commit

 

History

History
218 lines (167 loc) · 5.4 KB

GEMINI_SUPPORT.md

File metadata and controls

218 lines (167 loc) · 5.4 KB

Gemini CLI Runtime Support

Trinity now supports multiple AI runtimes, allowing you to choose between Claude Code (Anthropic) and Gemini CLI (Google) for your agents.

Why Multi-Runtime?

Cost Optimization:

  • Gemini: Free tier (60 req/min, 1,000/day)
  • Claude: Pay-per-use

Context Window:

  • Gemini: 1 million tokens (5x larger)
  • Claude: 200K tokens

Use Case Matching:

  • Gemini: Data processing, monitoring, large codebases
  • Claude: Complex reasoning, code quality, reliability

Getting Started

1. Get a Google API Key

  1. Go to Google AI Studio
  2. Click "Create API Key"
  3. Copy your key

2. Configure Trinity

Add to your .env file:

GOOGLE_API_KEY=your-google-api-key-here

3. Create a Gemini Agent

Via UI:

  1. Click "Create Agent"
  2. Enter agent name
  3. Select runtime: Gemini CLI
  4. Choose model: gemini-2.5-pro
  5. Click "Create"

Via API:

curl -X POST http://localhost:8000/api/agents \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-gemini-agent",
    "runtime": "gemini-cli",
    "runtime_model": "gemini-2.5-pro",
    "resources": {"cpu": "2", "memory": "2g"}
  }'

Via Template:

# config/agent-templates/my-agent/template.yaml
name: my-gemini-agent
runtime:
  type: gemini-cli
  model: gemini-2.5-pro
resources:
  memory: "2g"  # Gemini is lighter

Feature Parity Matrix

Feature Claude Code Gemini CLI Status
Chat Interface ✅ ✅ Identical
MCP Tools ✅ ✅ Supported
Cost Tracking ✅ ✅ Unified
Token Tracking ✅ ✅ Unified
Session Continuity ✅ ✅ --resume
Tool Restrictions ✅ ✅ --allowed-tools
YOLO Mode ✅ ✅ --yolo
Context Window 200K 1M 5x larger
Pricing Pay-per-use Free tier Different
Native Search ❌ ✅ Gemini only

Cost Comparison

Example: 100 Messages/Day

Claude Code (Sonnet 4.5):

  • Input: $3/M tokens
  • Output: $15/M tokens
  • Estimated: $5-10/day

Gemini CLI (2.5 Pro):

  • Free tier: 1,000 requests/day
  • Estimated: $0/day (within limits)

Architecture

Trinity uses a Runtime Adapter pattern:

User Chat → Backend → Agent Container → [Runtime Adapter]
                                              ↓
                                    ┌─────────┴─────────┐
                                    ↓                   ↓
                            ClaudeCodeRuntime    GeminiRuntime
                                    ↓                   ↓
                              claude CLI          gemini CLI

Both runtimes implement the same interface:

  • execute(prompt, model, continue_session)
  • configure_mcp(servers)
  • Return same format: (response, execution_log, metadata)

Instruction File (CLAUDE.md)

Both Claude Code and Gemini CLI read agent instructions from CLAUDE.md.

Why keep the name CLAUDE.md?

  • Backward compatibility with existing agents
  • Both runtimes understand markdown instruction files
  • Renaming would break existing templates
workspace/
├── CLAUDE.md           # ← Both runtimes read this
├── .mcp.json           # ← Claude Code only
└── ...

In your CLAUDE.md, you can write instructions that work for both:

# Agent Instructions

You are a helpful agent. [Instructions work for both Claude and Gemini]

## Tools Available
- filesystem operations
- web search
- etc.

MCP Configuration

Claude Code: Uses .mcp.json file

{
  "mcpServers": {
    "trinity": {
      "url": "http://mcp-server:8080/mcp",
      "headers": {"Authorization": "Bearer KEY"}
    }
  }
}

Gemini CLI: Uses CLI commands (Trinity handles this automatically)

gemini mcp add trinity http://mcp-server:8080/mcp

Trinity translates MCP configuration to each runtime's format automatically.

Switching Runtimes

You can run both Claude and Gemini agents simultaneously:

# Multi-runtime system
agents:
  orchestrator:
    runtime: claude-code  # Complex reasoning
    model: sonnet-4.5
    resources: {memory: "4g"}

  worker-1:
    runtime: gemini-cli  # Data processing
    model: gemini-2.5-pro
    resources: {memory: "2g"}

  worker-2:
    runtime: gemini-cli  # Monitoring
    model: gemini-2.5-pro
    resources: {memory: "2g"}

Cost: $2-3/day for orchestrator, $0 for workers

Troubleshooting

"Gemini CLI is not available"

  • Rebuild base image: ./scripts/deploy/build-base-image.sh
  • Gemini CLI is installed during image build

"GOOGLE_API_KEY not configured"

  • Add to .env: GOOGLE_API_KEY=your-key
  • Restart agent container

MCP Tools Not Working

  • Check gemini mcp list inside container
  • Verify Trinity MCP server is accessible
  • Check agent permissions

Limitations

Gemini Free Tier Limits:

  • 60 requests/minute
  • 1,000 requests/day
  • If exceeded, agent will fail with rate limit error

Workaround: Mix Claude and Gemini agents to stay within limits.

Next Steps

  1. Test: Create a test agent with local:test-gemini template
  2. Compare: Run same task on Claude and Gemini agents
  3. Optimize: Move simple tasks to Gemini, keep complex ones on Claude
  4. Monitor: Track costs via /api/observability/metrics