Ralph-On-Shelf (ROS)

Launch an autonomous Claude agent inside an E2B cloud sandbox with MCP connector support.

MCP Servers

When mcp=auto (default), Ralph analyzes your prompt and automatically connects to relevant MCP servers:

• exa-search: Web search and research (keywords: search, find, research, web)
• aws-knowledge: AWS documentation and cloud knowledge (keywords: aws, cloud, lambda, s3)

Use mcp=none to disable MCP and use the classic SDK mode. Use mcp=exa-search,aws-knowledge to explicitly specify servers.

Execute

Run the following Python code to start the Ralph loop:

cd {{cwd}} && python -c "
from main import launch_ralph

result = launch_ralph(
    prompt='''{{prompt}}''',
    max_iterations={{max-iterations | default: 10}},
    completion_promise='{{completion-promise | default: COMPLETE}}',
    output_dir='{{output-dir | default: ralph_output}}',
    mcp='{{mcp | default: auto}}',
    timeout=600
)

print(f'[ralph] final status: {result[\"status\"]}')
print(f'[ralph] total iterations: {result[\"iterations\"]}/{result[\"max_iterations\"]}')
print(f'[ralph] MCP servers: {result.get(\"mcp_servers\", [])}')
if result.get('downloaded_files'):
    print(f'[ralph] files saved to: {result[\"output_dir\"]}')
    for f in result['downloaded_files']:
        print(f'  - {f}')
"

The agent will run autonomously in the E2B sandbox until:

• The completion promise appears in the output
• Max iterations is reached