Agent Runner

Deploy Claude agents as MCP servers on Google Cloud Run with A2A communication, session reflection hooks, and auto-registration with Claude Code.

Quick Start

Prerequisites

• gcloud CLI authenticated (gcloud auth login)
• podman for container builds
• python3 and openssl
• ANTHROPIC_API_KEY env var set

One-Command Deploy

make bootstrap

This sets up GCP infrastructure, builds the container, deploys to Cloud Run, and prints your MCP server URL and OAuth credentials.

Connect to Claude Code

make connect          # Register with Claude Code CLI
make connect-oauth    # Print OAuth credentials for Claude.ai web
make mcp-json         # Generate .mcp.json for project-level auto-connect

Architecture

Each deployed agent is simultaneously:

• An MCP server (callers invoke tools via FastMCP Streamable HTTP)
• An MCP client (consumes external MCP servers defined in YAML config)
• An A2A server (serves Agent Card, accepts A2A tasks)
• An A2A client (discovers and calls remote agents)

Stack

Component	Package
Agent runtime	claude-agent-sdk
MCP server	fastmcp (PrefectHQ)
A2A protocol	a2a-sdk (Google)
Auth	PyJWT (OAuth 2.1, stateless JWTs)
Config	pydantic + pyyaml

Configuration

Edit agent-config.yaml or use env vars:

agent:
  name: "my-agent"
  model: "claude-sonnet-4-6"
  max_turns: 50
  timeout: 300

See agent-config.example.yaml for full documentation.

Development

make build                          # Build container image
make run-server                     # Start MCP server locally
make run-agent TASK="list buckets"  # One-shot CLI task
make test                           # Run tests
make lint                           # Run linter

Deploy a Custom Agent

# Edit agent-config.yaml with your agent settings
make build
make push
make deploy
make configure-url    # First deploy only
make connect          # Register with Claude Code

Credential Management

make show-credentials   # Display OAuth client ID/secret
make rotate-oauth       # Regenerate credentials (requires re-deploy)