Arcade agent templates

CLI to scaffold agents from Jinja2 templates and JSON configs.

All generated agents use Arcade as the tool runtime. Arcade handles auth and execution for external service integrations (Gmail, Slack, GitHub, etc.), so agents get tool access without managing OAuth flows or API keys directly.

Structure

templates/             # Jinja2 templates (ts_langchain, python_openai_agents_sdk, etc.)
agent-configurations/  # JSON configs organized by template type
real_agents/           # Generated output

Usage

python create_agent.py <config_path> [--template <name>] [--output-dir <path>]

Template is inferred from config's parent directory if not specified.

The --output-dir flag allows you to specify a custom output directory. If not provided, the agent will be created at real_agents/<template>/<config_stem>/ where <config_stem> is the config filename without extension.

Example

# Generate a ts_langchain agent from a config
python create_agent.py agent-configurations/ts_langchain/gmail-slack.json

This reads gmail-slack.json, applies values to the ts_langchain template, and outputs to real_agents/ts_langchain/gmail-slack/.

Config Format

{
  "arcade_tool_list": ["Gmail_ListEmails", "Gmail_SendEmail"],
  "arcade_toolkit_list": ["Slack", "gmail"],
  "agent_instruction": "Your system prompt here",
  "agent_description": "Short description"
}

Config keys become Jinja2 variables in templates.