Citadel Saas Factory

Clone. Set one API key. Run any of 265 agents. Ship to any server.

[Quick Start](#-quick-start) | [Agent Fleet](#-the-265-agent-fleet) | [Architecture](#-architecture) | [Runtime Engine](#-runtime-engine) | [Free Toolchain](#-free-toolchain)

---

Why This Exists

Most SaaS frameworks give you scaffolding. This one gives you a full autonomous workforce.

265 AI agents across 15 business domains -- from CEO strategy to DevOps canary deploys -- each with its own system prompt, RAG retrieval pipeline, model tier, and guardrails. The entire stack runs on any Linux server with Docker. No AWS. No vendor lock-in. No monthly SaaS bills.

You:     "Analyze our API for security vulnerabilities"
Citadel: Routes to sec-sast agent (reasoning_deep tier, Claude Opus)
         -> Retrieves codebase context via RAG
         -> Runs OWASP Top 10 analysis
         -> Returns grounded findings with citations
         -> Output validated through triple-layer guardrails

---

Quick Start

git clone https://github.com/Citadel-Cloud-Management/citadel-saas-factory.git
cd citadel-saas-factory
cp .env.example .env          # Add your ANTHROPIC_API_KEY
docker compose up -d           # Postgres, Redis, Keycloak, MinIO, RabbitMQ
cd backbone && pip install -e ".[dev]" && cd ..

Run your first agent in 3 lines:

from backbone.runtime.agent_sdk import AgentSDK
sdk = AgentSDK()
result = await sdk.run_agent("eng-api-designer", prompt="Design a health check endpoint")

Or use the CLI:

python -m backbone.cli run "Analyze competitor pricing strategy"
python -m backbone.cli health
python -m backbone.cli agents

---

The 265-Agent Fleet

Every agent has: a system prompt, RAG retrieval pipeline, model tier routing, and guardrails validation.

sdk.list_agents(domain="engineering")   # 25 agents
sdk.list_agents(domain="security")      # 22 agents
sdk.list_agents(domain="devops")        # 28 agents
sdk.list_agents()                       # all 265

Executive and Strategy 12

CEO Strategist, COO Operations, CFO Finance, CTO Technology, CMO Marketing, CPO Product, VP Engineering, VP Sales, OKR Tracker, Board Reporter, Competitive Intel, Decision Logger

Marketing and Growth 22

SEO, Content, Social Media, Email, PPC, Analytics, Landing Page, Brand Voice, PR, Influencer, Video, Podcast, Community, Growth Hacker, A/B Testing, Competitor Monitor, Newsletter, Webinar, Affiliate, Product Launch, Persona, Retention

Sales and Revenue 18

Lead Qualifier, Outbound, Proposals, CRM, Demo Prep, Contracts, Forecast, Win/Loss, Pricing, Territory, Upsell, Scheduler, Objection Handler, Referral, Partner, Call Analyzer, Pipeline Cleaner, Commission

Customer Success 15

Onboarding, Ticket Router, Response Drafter, Escalation, Churn Predictor, Health Scorer, NPS, Knowledge Builder, Chatbot, Feedback, Renewal, QBR, Adoption, SLA Monitor, Voice of Customer

Product and Design 20

UI Designer, UX Research, Wireframes, Prototyping, Design System, Accessibility, Responsive, Color Palette, Typography, Icons, Animation, Illustration, Data Viz, User Flow, Heuristic Eval, Onboarding UX, Form UX, Search UX, Notifications, Mobile

Engineering 25

API Designer, Models, Schemas, Services, Repositories, Migrations, Middleware, Events, Workers, Auth, Cache, Search, Webhooks, Email, File Handling, Pagination, Rate Limiting, Health, WebSockets, GraphQL, Multi-tenant, Error Handling, Logging, Config, Code Review

Frontend 18

Components, Pages, Layouts, Forms, Tables, Charts, Auth, State, API Client, Accessibility, Responsive, i18n, SEO, Performance, Testing, Animation, Error Boundaries, PWA

DevOps and Infrastructure 28

CI, CD, GitOps, Image Build, Image Scan, Image Sign, Helm, Terraform, Ansible, K8s, Scaling, Debugging, Canary, Rollback, Release, Certs, DNS, Backup, Restore, Monitoring, Alerts, Logs, Service Mesh, Ingress, Storage, Queue, Cost, Capacity

Security 22

SAST, SCA, DAST, Secrets, Container, IaC, Runtime, Policy, Incidents, Vulnerabilities, Patching, Access Review, RBAC, Encryption, Audit, Compliance, Pentest, Threat Hunt, Network, WAF, PII Detection, Supply Chain

Data and Analytics 18

Schema, Migrations, Indexes, Query Optimization, RLS, Backups, ETL, Analytics, Vectors, Warehouse, Reports, Dashboards, Events, Cohorts, A/B Analysis, Forecasting, Anomaly, Privacy

QA and Testing 22

Unit, Integration, E2E, API, Load, Performance, Security, Accessibility, Visual Regression, Coverage, Mutation, Contract, Chaos, Fixtures, Mocks, Regression, Flaky Detection, Prioritization, Smoke, Compatibility, Data Validation, Reporting

HR, Finance, Legal, Content 45

Job Descriptions, Interviews, Onboarding, Performance, Billing/Stripe, Subscriptions, Tax, Revenue, Runway, ToS, DPA, GDPR, SOC2, Tech Writing, Blog, Docs, Changelog, Case Studies, and 27 more

Full agent registry (click to expand)

See [.claude/agents/_registry.yaml](.claude/agents/_registry.yaml) for all 265 entries with IDs, domains, and descriptions.

Each agent definition lives in .claude/agents//.md with:

• YAML frontmatter (id, domain, tier, rag_enabled, tools, skills)
• RAG Retrieval Prompt with domain-scoped metadata filters
• LLM System Prompt with role, rules, and escalation protocol

---

Architecture

                              CLIENTS
                    Browser / Mobile / API
                              |
                    +---------+---------+
                    |                   |
              +-----v------+    +------v-------+
              |  FRONTEND  |    | AUTH GATEWAY  |
              | Next.js 14 |    | Keycloak 24   |
              +-----+------+    +------+-------+
                    |                  |
              +-----v------------------v-------+
              |          BACKEND API           |
              |   FastAPI + Agent Runtime      |
              |   /agent/run  /agent/browse    |
              +-----+------+------+------+-----+
                    |      |      |      |
            +-------v+  +-v----+ +v-----+ +v--------+
            |Postgres|  |Redis | |MinIO | |RabbitMQ  |
            |16+pgvec|  |  7   | | (S3) | |          |
            +--------+  +-----+ +------+ +---------+
                    |
            +-------v----------------------------+
            |        AGENT RUNTIME (backbone/)   |
            |  ModelRouter -> AgentSDK -> RAG    |
            |  Planner -> Guardrails -> Tracer   |
            +-------+---------------------------+
                    |
            +-------v----------------------------+
            |         ORCHESTRATION              |
            |   K3s + ArgoCD + Linkerd (mTLS)    |
            +------------------------------------+

Tech Stack

| Layer | Technology | Why | |-------|-----------|-----| | Backend | FastAPI (Python 3.12) | Async-first, Pydantic validation, OpenAPI | | Agent Runtime | backbone/ (LiteLLM + structlog) | Multi-model routing, budget control, RAG | | Frontend | Next.js 14 (TypeScript) | SSR/SSG, React Server Components | | Database | PostgreSQL 16 + pgvector | HNSW vector search, JSONB, RLS | | Cache | Redis 7 | Session, rate limiting, pub/sub | | Auth | Keycloak 24 | OAuth2, RBAC, MFA, SSO | | Storage | MinIO | S3-compatible, self-hosted | | Messaging | RabbitMQ | Async events, DLQ | | Orchestration | K3s + ArgoCD | Lightweight K8s, GitOps | | Service Mesh | Linkerd | mTLS between all services | | Secrets | HashiCorp Vault | Rotation, encryption | | Observability | Prometheus + Grafana + Loki | Metrics, dashboards, logs |

---

Runtime Engine

The backbone/ package powers the agent fleet with 5 model tiers, agentic RAG, and triple-layer guardrails.

Model Routing (5 tiers, automatic fallback)

reasoning_deep:   Claude Opus 4     -> Sonnet 4       # Architecture, security
reasoning_fast:   Claude Sonnet 4   -> Haiku 4.5      # Development, reviews
cheap_fast:       Claude Haiku 4.5  -> --              # Content, high volume
rag_specialist:   Claude Sonnet 4   -> Haiku 4.5      # RAG with citations
local_only:       Ollama Llama 3.1  -> Mistral         # Budget fallback

Daily budget cap ($50 default). Exceeds? Auto-routes to local_only. Service continues, quality degrades gracefully.

Agent SDK

from backbone.runtime.agent_sdk import AgentSDK

sdk = AgentSDK()

# Run any agent by ID
result = await sdk.run_agent("sec-sast", prompt="Scan auth module for injection")

# Spawn sub-agents (depth-capped at 3, fan-out at 5)
await sdk.spawn_sub_agent("parent-id", "sec-pentest", prompt="Deep scan /api/auth")

# List by domain
sdk.list_agents(domain="security")  # 22 agents

Agentic RAG Pipeline

Query -> Retrieve (vector + keyword + RRF) -> Grade (LLM-as-judge)
                                                |
                                    relevance  Executor runs -> Critic approves or revises

Triple-Layer Guardrails

| Layer | What | How | |-------|------|-----| | 1. Output Validation | Hallucination, toxicity, PII | Guardrails AI Hub validators | | 2. Input Sanitization | Prompt injection in RAG docs and browser content | 16 regex patterns, truncation | | 3. Action Enforcement | Out-of-scope tool calls | 10-action allowlist, fail-closed |

Threshold: 0.85. Below? Retry (max 3). Still below? Reject. No exceptions.

Chrome Extension

Manifest V3 side-panel assistant at extensions/chrome/. Captures active tab text on explicit user click, sanitizes through injection filters, routes through POST /agent/browse.

---

Harness (.claude/)

The full agent harness lives in .claude/:

| Component | Count | Purpose | |-----------|-------|---------| | Agents | 265 generated + 11 hand-written | Executable definitions with RAG prompts | | Skills | 10 | Deep reference: code-review, deploy, guardrails, llm-wiki, testing, etc. | | Commands | 38 (20 YAML + 18 MD) | /deploy, /audit, /scaffold, /wiki-ingest, etc. | | Hooks | 11 scripts + settings.json | pre-commit, post-deploy, on-security-alert, vault-autolink | | Templates | 20 | FastAPI endpoints, React components, Helm charts, Terraform modules | | Rules | 21 | Architecture, security, testing, naming, guardrails, obsidian backlinks | | MCP Servers | 6 | GitHub, filesystem, PostgreSQL, Docker, Kubernetes, Ruflo |

---

Free Toolchain

Total monthly software cost: $0

Orchestration

• ArgoCD (replaces Spinnaker)
• K3s (replaces EKS/GKE)
• Traefik (replaces AWS ALB)
• Linkerd (replaces Istio)

Security

• Semgrep (replaces SonarQube)
• Trivy (replaces Snyk)
• Falco (replaces Sysdig)
• Kyverno (replaces OPA GK)
• OWASP ZAP (replaces Burp)
• TruffleHog (replaces GitGuardian)

Identity and Secrets

• Keycloak (replaces Auth0)
• HashiCorp Vault (replaces AWS SM)
• Certbot (replaces commercial TLS)

Observability

• Prometheus (replaces Datadog)
• Grafana (replaces Datadog dashboards)
• Loki (replaces Splunk)
• Grafana OnCall (replaces PagerDuty)

Storage and Messaging

• MinIO (replaces AWS S3)
• RabbitMQ (replaces SQS)
• Velero (replaces Kasten K10)

Deployment

• Ansible (replaces Puppet/Chef)
• Flagsmith (replaces LaunchDarkly)

AI/ML Security

• Guardrails AI
• DeepEval

---

Docker Compose (One Command)

docker compose up -d

| Service | Port | Purpose | |---------|------|---------| | PostgreSQL 16 | 5432 | Primary database + pgvector | | Redis 7 | 6379 | Cache, sessions, rate limiting | | Keycloak 24 | 8080 | Auth server (admin console) | | MinIO | 9000/9001 | Object storage and console | | RabbitMQ | 5672/15672 | Message broker and management | | MailHog | 1025/8025 | Local email capture | | Backend | 8000 | FastAPI API | | Frontend | 3000 | Next.js UI |

---

Infrastructure Agnostic

Runs on any Linux server with SSH and Docker. No cloud vendor lock-in.

• Any VPS (Hetzner, DigitalOcean, Linode, Vultr)
• Bare metal servers
• On-premises infrastructure
• Edge deployments
• Home lab

---

LLM Wiki (Karpathy Pattern)

The agent fleet uses Andrej Karpathy's LLM Wiki pattern as persistent brain memory:

docs/vault/
  raw/          # Immutable source documents (never modified by LLM)
  wiki/         # LLM-maintained compiled knowledge
    index.md    # Master index (always consulted first)
    entities/   # People, tools, services
    concepts/   # Patterns, principles
    sources/    # Source metadata
  SCHEMA.md     # Governance (co-evolved with human)

Every session adds to the wiki. Every valuable answer is filed back. Knowledge compounds across all 265 agents instead of dying with the chat.

---

Security

• OWASP LLM Top 10 mapped: security/ai-security-by-design.html
• Sigma detection rules for injection attempts and repeated rejections
• OPA policies for browse endpoint rate limiting
• Kyverno policies for agent container restrictions
• All outputs validated. All inputs sanitized. Fail-closed by default.

---

Contributing

1. Fork the repo
2. Create a feature branch: git checkout -b feat/your-feature
3. Write tests first (TDD mandatory, 80% coverage minimum)
4. Run: make lint && make test && make security
5. Commit: feat: add user auth
6. Open a PR, wait for CI, get 1 approval

---

License

MIT License. See [LICENSE](LICENSE).

Copyright (c) Citadel Cloud Management

---

citadelcloudmanagement.com

Built with Claude Code. Powered by 265 autonomous agents. $0/month software cost.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Citadel-Cloud-Management
• Source: Citadel-Cloud-Management/citadel-saas-factory
• License: MIT
• Homepage: https://citadelcloudmanagement.com

Install and usage instructions live in the source repository linked above.