Developer Tools & Agent Integration

MCP tools

The framework exposes 12 MCP tools via hermes-redteam/mcp-server.ts for use with any MCP-compatible agent:

Tool	Purpose
read_repo	Scan source code for tools, roles, guardrails, secrets
probe_target	Send benign test message to observe API shape
read_prior_reports	Load previous scan results for adaptive planning
write_config	Generate and save a red-team config
write_custom_attacks	Create custom attack CSV/JSON files
write_policy	Create judge policy with per-category overrides
run_scan	Start a red-team scan via dashboard API
check_run_status	Poll scan progress (attacks completed, phase)
get_run_results	Get full results with verdicts and findings
cancel_run	Cancel a running scan
list_categories_and_strategies	List all 141 categories + 155 strategies with compliance mappings
suggest_guardrails	Map vulnerabilities to Votal Shield guardrail configs

Register with Hermes:

hermes mcp add wb-redteam -- npx tsx $(pwd)/hermes-redteam/mcp-server.ts

Natural conversation with Hermes:

You:    test my chatbot at http://localhost:3000 for safety issues
Hermes: [calls probe_target, write_config, run_scan]
        Scan started. 15 categories, 22 strategies, 2 rounds.

You:    show results
Hermes: [calls get_run_results]
        5 vulnerabilities found. 3 critical prompt injection, 2 high data exfiltration.

You:    how do I fix these?
Hermes: [calls suggest_guardrails]
        Deploy Votal Shield with adversarial-prompt-detection enabled.

AI Assistant

Standalone, no Hermes needed:

npm run ai

Natural-language terminal interface with intent classification, LLM-powered config generation, and Votal Shield guardrail recommendations.

Guardrail recommendations (Votal Shield)

When vulnerabilities are found, the framework maps them to specific Votal Shield guardrail configurations:

Vulnerability	Shield Guardrail	Endpoint
Prompt injection	adversarial-prompt-detection	/guardrails/input
Toxic content	toxicity-detection	/guardrails/output
PII disclosure	pii-detection + output-redaction	/guardrails/output
Hallucination	hallucination-detection	/guardrails/output
Data exfiltration	pii-detection + keyword-blocklist	/guardrails/output
Tool misuse	agentic tool authorization	/guardrails/output
Content filter bypass	keyword-blocklist + adversarial-prompt-detection	/guardrails/input
Harmful advice	topic-restriction + toxicity-detection	/guardrails/input + output

Deploy Shield as a proxy — no code changes needed:

/v1/shield/chat/completions   # instead of /v1/chat/completions