Themis · Aegis

Platform-agnostic AI skill compiler and multi-agent threat analysis — built for defenders.

Two subsystems, one deployment:

| | Aegis | Themis | |---|---|---| | What | Skill compiler + marketplace | LangGraph orchestrator | | Does | Write once → deploy to any AI platform | Fan-out parallel threat analysis | | Format | SKILL.md → system prompt / ChatGPT Action / MCP | POST request → structured findings report |

Aegis

Author defensive security skills in a single SKILL.md file. Aegis compiles it to every platform format — one source, deployed everywhere.

Skill library

| Skill | Framework | Purpose | |---|---|---| | mitre-attack | ATT&CK | Adversary TTP mapping and detection guidance | | mitre-engage | Engage | Deception activity planning and execution | | deception-engineering | Engage + ATT&CK | Honeypot and honeytoken deployment | | mitre-atlas | ATLAS | AI/ML attack surface defence | | threat-modeling | STRIDE | Systematic threat modelling for systems | | threat-hunting | — | Hypothesis-driven hunt methodology | | network-security | — | Network segmentation and monitoring | | endpoint-security | — | Endpoint detection and hardening | | security-operations | — | SOC workflow and incident response | | attack-surface-mapping | — | External and internal ASM | | vulnerability-management | — | CVE triage and remediation prioritisation | | compliance-frameworks | — | NIST, ISO 27001, CIS Controls mapping | | supply-chain-security | — | Dependency and build pipeline hardening |

Quick start

# Install the compiler
npm install -g @aegis/compiler

# Compile a skill
aegis compile skills/mitre-attack --base-url https://your-deployment.vercel.app

# Deploy: push to Vercel — the marketplace and API are live automatically

Deploying a skill

After compilation, three artifacts are generated in skills/<name>/artifacts/:

• system-prompt.txt — paste into Claude Projects, Gemini Gems, or any chat UI
• openai-action.json — import into ChatGPT GPT Builder → Add Action
• mcp-manifest.json — wire into claude_desktop_config.json for Claude Desktop / Cursor

Themis

An AI-powered threat analysis engine. Themis accepts a task description and security context, decomposes it into sub-tasks, fans out to specialist skill agents in parallel, applies guardrails to every output, and synthesises a structured findings report.

Architecture

Task input
    │
  specNode          — decompose into sub-tasks
    │
  fanOutNode        — Send() parallel dispatch
    │
  ┌──────────────────────────┐
  │ skillAgentNode (×N)      │  — ReAct agents, one per sub-task
  │ + guardrailNode          │  — block / flag / pass
  └──────────────────────────┘
    │
  synthesisNode     — reduce to structured report
    │
  auditNode         — write metadata to SQLite (no findings stored)

LangGraph MemorySaver holds graph state in-RAM for the duration of a session. Nothing persists beyond the request lifecycle except a metadata-only SQLite debrief row.

API

curl -X POST https://your-deployment.vercel.app/api/themis \
  -H 'Content-Type: application/json' \
  -d '{
    "task": "Assess the attack surface for a hybrid cloud + OT environment",
    "context": {
      "environments": ["enterprise", "hybrid", "ot"],
      "attackSurfaceTags": ["network", "lateral-movement", "credential-theft"]
    }
  }'

Response:

{
  "report": "## Findings\n...",
  "guardrailSummary": { "passed": 3, "flagged": 0, "blocked": 0 },
  "skillTrace": ["mitre-attack", "deception-engineering"],
  "totalInputTokens": 12400,
  "totalOutputTokens": 3200,
  "durationMs": 8400,
  "threadId": "abc-123"
}

SSE streaming — pass Accept: text/event-stream to receive node-name events as the graph executes. Node content is never included in stream events.

Requirements

At least one LLM provider API key:

ANTHROPIC_API_KEY=...      # Claude — preferred
OPENAI_API_KEY=...         # GPT-4o-mini — fallback
GOOGLE_GENERATIVE_AI_API_KEY=...  # Gemini — fallback

LangSmith tracing (optional):

LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=...
LANGCHAIN_PROJECT=themis

Security

Key constraints enforced in the codebase:

• LLM SDKs server-only. Provider SDKs are importable only in lib/themis/provider.ts. Next.js serverExternalPackages enforces this at bundle level.
• No findings persistence. Task content and LLM outputs are never written to any store — not SQLite, not logs, not external services.
• Sanitised errors. All client-facing errors pass through safeError() — no stack traces, paths, or model names reach the client.
• Sanitised logs. Loggers receive only metadata (hashes, token counts, durations, skill slugs). No user input or LLM output appears in logs.
• Path traversal prevention. Skill name and phase ID are validated against an allowlist before any file read.

See docs/TECHNICAL.md for the full security model and data flow.

Development

npm install
cp .env.local.example .env.local   # add at least one LLM key
npm run dev                         # http://localhost:3000
npm test                            # 310 tests
npm run build                       # production build

Project structure

├── app/api/themis/        # Themis streaming route (App Router)
├── bin/                   # aegis CLI
├── components/            # React UI components
├── docs/                  # Technical documentation
│   └── TECHNICAL.md
├── lib/
│   ├── skill-reader.js    # Aegis skill loader with path traversal protection
│   └── themis/            # Themis engine
│       ├── graph/         # LangGraph nodes, state, tools
│       ├── debrief.ts     # SQLite audit writer
│       ├── llm-factory.ts # Provider selection
│       ├── provider.ts    # LLM SDK (server-only)
│       └── index.ts       # orchestrate() entry point
├── pages/                 # Next.js pages router (Aegis UI + API)
├── skills/                # Skill source files + compiled artifacts
└── styles/globals.css     # Design system

License

MIT © Drupad Sachania