Agent Specialists (The Workforce)¶
Emmu is powered by an extensive library of specialized AI personas. Rather than relying on one massive, generalized prompt, Emmu utilizes the AgentFactory to spin up highly focused domain experts.
[!NOTE] The Problem: Supplying a single LLM with tools for Web Exploitation, Binary Analysis, Cloud Auditing, and Mobile Pentesting simultaneously blows out the Context Window and causes severe model confusion.
The Solution: The
SpecialistRegistrydivides knowledge into 60+ micro-agents. The Orchestrator dynamically loads only the specific agent needed for the immediate task, injecting it with targeted tools and a precise System Prompt.
1. The Specialist Categories¶
Emmu's specialists.py registry maintains a roster of agents divided into six distinct operational groups:
- Web Vulnerabilities (20+):
sqli,rce,xss,proto_pollution,jwt_expert, etc. - Infrastructure & Cloud (10+):
aws_expert,k8s_expert,active_directory, etc. - Advanced Recon & Fuzzing (10+):
osint_pro,subdomain_king,mobile_recon, etc. - Binary & Reverse Engineering (10+):
bof_expert,rop_master,re_binary,malware_analyst, etc. - CTF Specific (10+):
ctf_crypto,ctf_pwn,ctf_strategy, etc. - Special Purpose: The
reviewer(Anti-hallucination) andmission_coordinator.
2. The [BRIDGE PARITY] Prompt Injection¶
For an AI to act autonomously, it must understand that it possesses physical agency.
When the AgentFactory instantiates an agent, it dynamically injects the [BRIDGE PARITY] instruction block into the system prompt. This block explicitly teaches the LLM that it is not merely a chat bot, but an in-process bridge with direct access to local functions.
Key capabilities granted via Bridge Parity:
- Reading and patching solve.py scripts.
- Executing scripts via run_ctf_solver_command.
- Interacting with Android devices via mobile_adb_command.
- Saving validated findings to the dashboard using update_emmuoec_finding.
3. Think-Stripping for Reasoning Models¶
With the advent of deep-reasoning models (like DeepSeek R1), a new problem emerged in AI orchestration: models outputting massive <think> blocks that corrupt JSON function calls and clog terminal UI.
The Solution: Emmu implements ThinkStrippingClient, a specialized wrapper over the AsyncOpenAI client.
# Emmu automatically strips <think> blocks before they reach the tool parser
choice.message.content = re.sub(
r"<think>.*?</think>",
"",
choice.message.content,
flags=re.DOTALL
).strip()
use_think_stripping is enabled (automatically done for local DeepSeek/Qwen instances via LM Studio), Emmu strips the raw reasoning from the response. This allows the model to "think" deeply on the backend, while Emmu only parses the final, clean, actionable output for tool execution.
4. LLM Providers & Multi-Model Support¶
Emmu is model-agnostic. While it is optimized for high-reasoning models, the AgentFactory can orchestrate a diverse fleet of LLM providers simultaneously.
Supported Cloud Providers¶
- OpenAI: Full support for GPT-4o, GPT-4-turbo, and GPT-3.5.
- Groq (Ultra-Speed): Native integration for Llama 3, Mixtral, and Gemma via the Groq Cloud API. Recommended for rapid-response specialists (e.g.,
recon). - OpenRouter (The Gateway): Access to Claude 3.5 Sonnet, Grok-1, and Codex models through a single API key.
- Gemini API: Support for Gemini 1.5 Pro and Flash via OpenAI-compatible endpoints.
Local & Offline Models¶
Emmu excels at orchestrating local models for sensitive missions or offline CTF environments:
- Ollama: Connect natively to your Ollama instances.
- LM Studio: Seamlessly bridge to local servers hosting DeepSeek R1, Qwen 2.5, or Llama 3.
- Self-Hosted: Any model served via an OpenAI-compatible API (v1/completions) can be integrated.
5. Custom "OpenAI-Compatible" Agents¶
You can define your own agent profiles in the config.yaml or environment variables. This allows you to add any new model that hits the market as long as it supports the OpenAI standard.
Example Configuration Profile:
local_agents:
- id: "my-custom-deepseek"
name: "DeepSeek-R1-Local"
service: "lmstudio"
base_url: "http://192.168.1.10:1234/v1"
model: "deepseek-r1-distill-llama-70b"
use_think_stripping: true
enabled: true
Next Steps¶
Now that you understand the agents, learn how Emmu prevents them from forgetting their mission history in the State Persistence & Memory guide.