gemini-adk-cli-rs

Command-line interface for ADK agent development. Mirrors the upstream adk CLI — scaffold, run, serve, evaluate, and deploy agents.

Prerequisites

• Rust toolchain (stable, 1.75+)
• Google API key or GCP credentials for agent execution
• Docker (for deploy cloud_run and deploy gke)
• gcloud CLI (for deploy cloud_run and deploy agent_engine)
• kubectl (for deploy gke)

Installation

# From workspace root
cargo install --path tools/gemini-adk-cli-rs

# Or run directly
cargo run -p gemini-adk-cli-rs -- <command>

The binary is named adk.

Commands

adk create — Scaffold a new agent

Creates a new agent project directory with boilerplate files.

adk create my_agent
adk create my_agent --model gemini-2.5-pro
adk create my_agent --api-key sk-abc123

Generated files:

my_agent/
├── agent.toml       # Agent configuration
├── Cargo.toml       # Rust project manifest
├── src/main.rs      # Entry point
├── .env             # API key
└── .gitignore

agent.toml format (TOML):

name = "my_agent"
description = "A new ADK agent"
model = "gemini-2.0-flash"
instruction = "You are a helpful assistant."
tools = []
sub_agents = []

adk run — Interactive terminal REPL

Run an agent interactively in the terminal.

# Basic usage — point at agent directory
adk run my_agent/

# Resume a session
adk run my_agent/ --session-id "session-abc-123"

# Save transcript on exit
adk run my_agent/ --save-session transcript.json

# Replay a saved session (non-interactive)
adk run my_agent/ --replay transcript.json

REPL commands:

• Type messages normally to chat with the agent
• /quit or /exit — exit the REPL
• Ctrl+D (EOF) — exit the REPL

Saved session format:

[
  { "user": "What's the weather in Tokyo?", "agent": "The weather in Tokyo is..." },
  { "user": "How about NYC?", "agent": "In New York City..." }
]

adk web — Development web server with UI

Start the web UI for interactive agent development.

adk web my_agent/
adk web my_agent/ --port 9000
adk web my_agent/ --host 0.0.0.0 --reload
adk web my_agent/ --a2a --trace-to-cloud

Flags:

adk api — Headless API server

Start a REST API server without the web UI. Same flags as adk web.

adk api my_agent/
adk api my_agent/ --port 8080 --allow-origins "http://localhost:3000"

Endpoints served:

GET  /list-apps                                    — List discovered agents
POST /run                                          — Execute agent (JSON)
POST /run_sse                                      — Execute agent (SSE stream)
GET  /apps/:app/users/:user/sessions/:session      — Get session
POST /apps/:app/users/:user/sessions               — Create session
DELETE /apps/:app/users/:user/sessions/:session     — Delete session
GET  /debug/trace/:event_id                        — Get trace spans
GET  /debug/trace/session/:session_id              — Get session traces

adk eval — Run evaluations

Evaluate an agent against a test set.

# Basic eval
adk eval my_agent/ tests/weather.evalset.json

# With scoring config and detailed output
adk eval my_agent/ tests/weather.evalset.json \
  --config-file tests/test_config.json \
  --print-detailed-results

.evalset.json format:

{
  "name": "Weather Agent Tests",
  "cases": [
    {
      "id": "tokyo_weather",
      "inputs": ["What is the weather in Tokyo?"],
      "expected": ["sunny", "temperature"],
      "tags": ["basic"]
    },
    {
      "id": "multi_turn",
      "inputs": [
        "What's the weather in NYC?",
        "How about tomorrow?"
      ],
      "expected": ["forecast"],
      "tags": ["multi-turn"]
    }
  ]
}

test_config.json format:

{
  "pass_threshold": 0.7,
  "criteria": [
    { "name": "response_match_score", "weight": 1.0 },
    { "name": "tool_trajectory_avg_score", "weight": 0.5, "description": "Tool usage accuracy" }
  ]
}

Available evaluation criteria:

adk deploy — Deploy to cloud

Deploy an agent to a cloud target.

# Cloud Run
adk deploy cloud_run my_agent/ --project my-gcp-project --region us-central1

# With web UI bundled
adk deploy cloud_run my_agent/ --project my-gcp-project --with-ui

# GKE
adk deploy gke my_agent/ --project my-gcp-project --service-name weather-svc

# Vertex AI Agent Engine
adk deploy agent_engine my_agent/ --project my-gcp-project

Cloud Run generates a Dockerfile and prints the gcloud run deploy command.

GKE generates a Dockerfile + k8s.yaml (Deployment + Service) and prints docker build / kubectl apply commands.

Agent Engine deploys to Vertex AI Agent Engine (API integration pending).

Flags:

Agent Discovery

The CLI discovers agents by looking for manifest files:

1. agent.toml — TOML format (used by adk create, adk run, adk eval, adk deploy)
2. agent.json / root_agent.json — JSON format (used by gemini-adk-api-rs, matches upstream ADK)

Both formats support the same fields. The CLI scans:

• <agent_dir>/agent.toml (single agent)
• <agent_dir>/<subdir>/agent.toml (multi-agent project with sub-agents in subdirectories)

Project Structure

tools/gemini-adk-cli-rs/
├── src/
│   ├── main.rs              — CLI entry point (clap derive)
│   ├── manifest.rs          — Agent manifest loading and discovery
│   └── commands/
│       ├── mod.rs           — Command module re-exports
│       ├── create.rs        — Project scaffolding
│       ├── run.rs           — Interactive REPL
│       ├── web.rs           — Web server with UI
│       ├── api.rs           — Headless API server
│       ├── eval.rs          — Evaluation runner
│       └── deploy.rs        — Cloud deployment
└── Cargo.toml