Arthor-Agent

Arthor-Agent (repository name DocSentinel) is an AI-driven SSDLC (Secure Software Development Lifecycle) platform for security teams, used to automate security activities across all stages of the software development lifecycle. It automatically reviews security-related documents, forms, and reports through AI agents orchestrated by LangGraph, compares inputs against policies and knowledge bases, and produces structured assessment reports containing risks, compliance gaps, and remediation recommendations.

1. Purpose and Positioning DocSentinel aims to embed security throughout the entire process from requirements to operations, rather than only reviewing documents before release. It is suitable for enterprises that need to scale security assessments across multiple projects and SSDLC stages without increasing headcount proportionally. The project consists of a React console, a FastAPI service, and an MCP/A2A agent gateway, and can be integrated into local security review workflows, CI/CD pipelines, AI agents, and multi-agent platforms, while not granting external agents approval authority.

2. Main Capabilities

3. Full SSDLC lifecycle coverage: Six dedicated AI agents, each with stage-specific skills, prompts, and knowledge base collections, can run a single stage or an end-to-end assessment—requirements (security requirements, compliance mapping, initial risk analysis), design (architecture review, STRIDE/DREAD threat modeling, SDR), development (secure coding standards, code review findings), testing (SAST/DAST report triage, penetration testing assessment), deployment (release readiness, configuration security, hardening), operations (incident response, vulnerability monitoring, log auditing).

4. Automated security assessment: After submitting security questionnaires, design documents, or audit reports, it uses the configured LLM to analyze and identify security risks (graded by severity), compliance gaps (against frameworks such as ISO 27001 and PCI DSS), and remediation steps.

5. Governance and compliance workflows: The policy package includes generic-ssdlc and eight public overlays (NIST SSDF, MAS TRM, ISO 27001:2022, EU AI Act, ISO 42001, China MLPS 2.0, OWASP SAMM, EU CRA); the Gate 1/3 questionnaire, submission, approval, audit, and evidence tracking processes are supported by SQLModel and Alembic migrations; framework overlays generate applicable SCD controls and expected evidence; Pallas Lens provides project readiness scoring.

6. LangGraph agent orchestration: Stateful, graph-based agent workflows with conditional branching by SSDLC stage; structured nodes are dividedFrom context collection, drafting, evidence review, pattern transformation, and governance persistence; supports cross-stage traceability, conditional routing, and human-in-the-loop (assessments created by agents enter the review queue and cannot approve themselves).

7. Threat Evidence Critic: STRIDE threats in the design phase are independently verified against the currently uploaded architecture document, giving a supported, contradicted, or insufficient_evidence determination, with a support score and rationale; evidence citations use content hashes and stable line locators, and display the exact source excerpt; only paragraphs from the current project document are accepted, verification failures safely abstain as insufficient_evidence, and all determinations still require human review.

8. RAG knowledge base: organizational security policies, standards, and historical audits can be uploaded; collections are divided by stage to ensure the most relevant context is retrieved.

9. Multi-format input and multi-LLM support: supports parsing PDF, Word, Excel, PPT, and text into a unified format for LLM use; multiple LLMs such as OpenAI, Claude, Qwen, or Ollama (local) can be used through a single interface.

10. Structured output: generates JSON/Markdown reports containing risk items, compliance gaps, and actionable remediation measures.

11. API-first, MCP- and A2A-ready: a local React console is used for human review, the REST API integrates with CI/CD, MCP exposes assessment capabilities to AI agents such as Claude Desktop, Cursor, and OpenClaw, and A2A-compatible platforms can delegate assessment tasks to DocSentinel.

12. Reproducible evaluation tooling: includes evals/ tools, with the first path covering the OWASP Benchmark v1.2 SAST category, supporting hard-key scoring and threat grounding scoring, and outputting machine-readable and human-readable scorecards.

III. How to Use Security personnel work in the React console; REST clients call /api/v1/*; coding agents call MCP or A2A through the agent gateway. REST write paths use JWT/RBAC dependencies, agent protocols use bearer gateway tokens or loopback-only development access, and high-LLM-cost POST paths are rate-limited by IP or bearer token. Assessment submissions enter AssessmentService, create an asynchronous task, parse uploaded or approved local documents, apply guardrails, and invoke the LangGraph assessment pipeline. The pipeline loads the selected skill, wraps untrusted document content as data, retrieves policy/history/evidence from the knowledge base,and then asks the LLM to generate draft/review text and convert it into a structured assessment schema. Deterministic services remain authoritative over governance decisions. The knowledge base persists chunks to Chroma, optionally graph artifacts to LightRAG, and reuses prior assessment history. Runtime LLM settings are loaded from .env and llm_config.json.

IV. Console Features Command Center: real-time API and LLM status, assessment throughput, review needs, remediation queue, and recent activity. Assessment Workbench: upload project documents, select SSDLC stage/skill, inspect AI-generated risks, and complete human review. Governance Portal: create projects, apply public framework overlays, generate controls, submit evidence, and track Pallas Lens readiness. Knowledge Base: ingest policies and standards for RAG review. Agent Integration: expose governed MCP and A2A tools to coding agents and multi-agent platforms without granting approval authority. Settings: switch providers such as DeepSeek, OpenAI, Anthropic, Qwen, or Ollama, with API keys accepted locally and only masked previews displayed.

V. Limitations and Boundaries Agents only make recommendations; deterministic policies and authorized personnel are responsible for acceptance. Human review is mandatory for agent gateway submissions, and external agents cannot approve their own work. There are MCP document root restrictions, token/loopback boundaries, knowledge base result boundaries, input guardrails, and LLM base URL network checks. Verification failures safely abstain as insufficient_evidence. Other specific details such as features, fees, pricing, platform support, or effectiveness are not yet provided.