░█████╗░███████╗░█████╗░██████╗░███████╗
  ██╔══██╗██╔════╝██╔══██╗██╔══██╗██╔════╝
  ██║░░╚═╝█████╗░░██║░░██║██████╦╝█████╗░░
  ██║░░██╗██╔══╝░░██║░░██║██╔══██╗██╔══╝░░
  ╚█████╔╝███████╗╚█████╔╝██████╦╝███████╗
  ░╚════╝░╚══════╝░╚════╝░╚═════╝░╚══════╝

Autonomous AI Engineering Orchestrator

v1.16.0 · Three Brains Architecture · Model-Agnostic · MIT

What is Ceobe?

Ceobe is a CLI-native AI Engineering Orchestrator that takes a description, image mockup, or PRD document — and produces a fully working codebase.

It is not a chat assistant. It is a pipeline.

  𝙿lanner → QA Auditor → 𝙴xecutor

Three independent AI brains, each with a separate provider, review each other's work before any code is written. The result is structured, architecture-compliant software — not hallucinated boilerplate.

Three Brains Architecture

| Brain | Role | Best Model | |---|---|---| | Planner | Generates BRD → Design → Architecture → Task Plan | Claude Sonnet, GPT-4o | | QA Auditor | Reviews blueprint against 18 engineering rules | Gemini 2.5 Flash, DeepSeek | | Executor | Writes the code from the validated plan | GLM-5.1, Kimi-K2, Llama 3 |

Using three different providers prevents self-evaluation bias and catches hallucinations early.

Installation

npm install -g ceobe-mastery-cli

Quick Start

# 1. Configure providers
ceobe setup

# 2. Build something
ceobe auto "Build a REST API with Go, Fiber, and PostgreSQL"

# 3. Add a feature to an existing project
ceobe auto --feature "Add Stripe payment gateway integration"

# 4. From a PRD document
ceobe auto --file requirements.md

# 5. From a UI mockup image
ceobe auto --file dashboard.png "Use Next.js + Tailwind"

Configuration

# API Keys
ceobe key set gemini         <GEMINI_API_KEY>
ceobe key set anthropic      <ANTHROPIC_API_KEY>
ceobe key set glm            <GLM_API_KEY>

# Assign roles to providers
ceobe key set planner-provider   anthropic
ceobe key set qa-provider        gemini
ceobe key set executor-provider  glm

# Optional: model overrides
ceobe key set planner-model  claude-4-5-sonnet
ceobe key set executor-model glm-5.1-flash

# Verify setup
ceobe doctor

Key storage priority (highest → lowest):

1. System environment variable (export KEY=...)
2. ~/.ceobe/keys.json (via ceobe key set)
3. .env file in project directory

AI Providers

| Slug | Provider | Default Model | |---|---|---| | gemini | Google Gemini | gemini-2.5-flash | | anthropic | Anthropic Claude | claude-4-5-sonnet | | glm | Zhipu AI | glm-5.1-flash | | kimi | Moonshot AI | kimi-k2.6-plus | | deepseek | DeepSeek | deepseek-v3 | | groq | Groq | llama-3.3-70b-versatile | | openai | OpenAI | gpt-4o | | qwen | Alibaba | qwen-3-max | | together | Together AI | llama-3.1-70B | | ollama | Local | llama3.2 (no key) | | custom | Any OpenAI-compatible | — |

For custom providers:

CEOBE_PLANNER_BASE_URL=https://my-api.example.com/v1
CEOBE_PLANNER_API_KEY=sk-...

Optional Cloudflare AI Gateway routing: set CLOUDFLARE_ACCOUNT_ID + CLOUDFLARE_GATEWAY_ID.

CLI Reference

Core

ceobe auto [desc]          # Full pipeline: plan → audit → execute
ceobe plan [desc]          # Generate plan only (no execution)
ceobe audit                # QA-review the generated plan against engineering rules
ceobe execute              # Execute an existing approved plan
ceobe export-rules         # Sync rules to Cursor / Windsurf / Cline / Antigravity

ceobe auto flags:

| Flag | Description | |---|---| | --ask | Human-in-the-loop — confirm before each destructive action | | --feature | Brownfield mode — add to existing codebase without breaking it | | --file <path> | Use a .md PRD or .png/.webp mockup as requirements | | --sandbox | Isolate AI execution inside a Docker container | | --worktree | Run in an isolated git worktree, merge on success | | --creative | Bypass Ponytail Ladder — allow over-engineered exploration |

Workspace

ceobe status               # Pipeline progress, phases & generated files
ceobe index                # Build semantic memory index (RAG) for workspace
ceobe trim                 # Whole-repo bloat scan — detect over-engineering
ceobe debt                 # Scan tech-debt markers (// ceobe: / // ponytail:)
ceobe reflect [--auto-skill] # Analyze logs, auto-generate skill draft
ceobe rollback             # Hard-reset to pre-AI git snapshot
ceobe reset --yes          # Clear all plans & state
ceobe recon <url>          # Dynamic reverse-engineering of a URL

Configuration

ceobe setup                # Interactive first-time setup wizard
ceobe key set <p> <val>    # Save API key or provider config
ceobe key list             # Show all keys & active providers
ceobe key get <provider>   # Peek at a key (masked)
ceobe key remove <provider> # Delete a stored key
ceobe mode [autonomous|ask] # View or change execution mode
ceobe doctor               # Diagnose keys, providers & workspace
ceobe cost                 # Live token usage & API cost report

Advanced

ceobe benchmark            # LLM benchmark — compare models on accuracy & cost
ceobe log [-n <lines>]     # View execution log
ceobe daemon --telegram    # Start remote Telegram bot
ceobe mcp                  # Launch MCP stdio server (for AI IDE integrations)
ceobe skill list           # Browse all 288 available skills
ceobe templates            # Manage project document templates

Execution Modes

Autonomous (default)

ceobe mode autonomous
ceobe auto "Build something"

Ceobe executes without interruption. Best for trusted, well-defined tasks.

Human-in-the-Loop (ask)

ceobe mode ask
# or per-run:
ceobe auto "Refactor auth module" --ask

Actions requiring confirmation in ask mode:

• write_file · edit_file · delete_file
• execute_command · rename_file · move_file
• start_background_service

Confirmations are delivered via readline in terminal or Telegram Inline Buttons when using the daemon.

Telegram Daemon

Control Ceobe remotely from your phone.

# 1. Set up credentials
ceobe key set telegram-token        <BOT_TOKEN>
ceobe key set telegram-allowed-users <YOUR_TELEGRAM_USER_ID>

# 2. Start the daemon
ceobe daemon --telegram

| Command | Action | |---|---| | /start | Wake bot & check status | | /projects | List registered workspaces | | /addproject <name> <path> | Register a workspace | | /cd <name> | Switch active workspace | | /mode <ask\|autonomous> | Toggle HITL / full-auto | | /auto | Shortcut — set mode to autonomous | | /ask | Shortcut — set mode to ask | | /status | Pipeline status + queue | | /cost | Live token usage & cost | | /logs | Tail last 50 execution lines | | /read <file> | Read a project file remotely | | /index | Build semantic memory index (RAG) | | /doctor | Run system diagnostics | | /reflect | AI self-reflection on logs | | /reset | Reset pipeline state | | /worktree | Toggle worktree isolation mode | | /cancel | Clear task queue | | /help | Show command list | | <any message> | Treated as a task instruction |

Safety & Security

Git Safety Net

• Auto-snapshot before every AI execution (git commit)
• Auto-rollback if self-healing fails after 3 retries (git reset --hard)
• Worktree isolation (--worktree) — merge to main only on success

Budget Guard

• Default limit: $5 USD per session (CEOBE_MAX_BUDGET)
• Hard-stop with BUDGET_EXCEEDED error before API call
• Custom pricing override via CEOBE_PRICING_OVERRIDE

Static Rule Checker

Runs automatically after execution on changed TypeScript files:

| Rule | Check | |---|---| | RULE-01 | Controllers must not import from /repository or /db | | RULE-08 | No direct database access (prisma., db.) in controllers | | RULE-13 | No hallucinated imports — all relative imports must resolve | | RULE-14 | Mandatory file header // Tujuan: in every file | | RULE-16 | No SELECT *, no N+1 queries in loops |

Prompt Injection Hardening

All user input is wrapped in <user_input> tags with an explicit system-level instruction to ignore any embedded jailbreak attempts.

Execution Limits

• Max 50 iterations per executor session (anti-infinite-loop)
• Max 3 self-heal cycles per session (JSON errors + command failures tracked separately)
• Token bleed protection — detects max_tokens exhaustion and recovers gracefully

Memory & RAG System

Ceobe builds a local semantic index of your workspace:

ceobe index   # Build once (or after major changes)

• Hybrid search: Reciprocal Rank Fusion (RRF) combining vector search + full-text BM25
• AST compression: TypeScript files are compressed to signature-only before embedding (~80% token savings)
• Incremental indexing: Only re-indexes files that changed (mtime cache)
• Dependency graph: Understands cross-module import relationships
• Tech-debt ledger: Tracks // ceobe: and // ponytail: markers

Supported embedding providers: gemini · openai · glm · ollama · together · custom

Plugin System

Drop a .ts file into .ceobe/plugins/ — it becomes a new tool available to the Executor immediately.

.ceobe/
  plugins/
    my-custom-tool.ts   ← exported { name, description, input_schema, execute() }

No core modifications needed. The plugin loader auto-discovers and injects tools per session.

Skills Library (288 skills)

Ceobe ships with 288 specialized skill files that are dynamically injected into prompts based on task context.

| Category | Count | Examples | |---|---|---| | Security / CTF | ~80 | ctf-pwn-heap, ctf-crypto-rsa, ctf-pwn-kernel | | Vulnerability | ~50 | vuln-sqli, vuln-jwt, vuln-ssrf, vuln-llm-attacks | | Recon | ~12 | recon-subdomain, recon-dorking, recon-secrets | | Payloads | 13 | payload-xss, payload-sqli, payload-xxe | | Post-Exploit | 7 | post-linux-privesc, post-container-escape | | Web Frameworks | 11 | framework-laravel, framework-nextjs, framework-django | | Mobile | 5 | flutter, react-native, mobile-android, mobile-ios | | Backend | 8 | golang, python-backend, elysia, hono | | Database | 4 | prisma-orm, drizzle, database-architect | | DevOps | 7 | tech-docker, tech-kubernetes, deployment-ops, tech-cicd | | Testing | 6 | vitest, bdd-testing, visual-qa |

ceobe skill list   # Browse all available skills

MCP Server

Expose Ceobe's capabilities to any MCP-compatible AI coding tool (Cursor, Claude Desktop, etc.):

ceobe mcp

Available MCP tools:

| Tool | Description | |---|---| | ceobe_extract_ast | Compact AST summary of workspace | | ceobe_scan_debt | Scan tech-debt markers | | ceobe_trim_codebase | Full-repo bloat analysis | | check_tool_status | Check if a security tool is installed | | run_pentest | Launch autonomous pentest session |

Available MCP prompt: ceobe-engineering-rules — injects Ceobe's 18 engineering rules into AI context.

Engineering Rules

Ceobe enforces 18 architecture rules across the planning and verification phases:

1. Layered Architecture — strict layer boundaries, Adaptive Legacy Fallback
2. Separation of Concerns — one responsibility per module
3. Thin Controllers — no business logic in HTTP handlers
4. Explicit Error Handling — no silent catches
5. Consistent Naming — enforced conventions per language
6. Modular Design — small, focused, reusable modules
7. Framework Isolation — core logic must not depend on framework internals
8. Repository Pattern — all DB access via repository layer
9. Simplicity Over Cleverness — YAGNI, boring tech wins
10. Decision Documentation — ADRs required for non-obvious choices
11. Task Decomposition — all work planned before execution
12. Safe Modification — "Add, Don't Break" strategy
13. AI Strict Constraints — no hallucinated imports, verify before editing
14. Mandatory File Header — // Tujuan: required in every file
15. Token Efficiency — compact prompts, read before edit
16. DB Performance — no SELECT *, no N+1 in loops
17. Pre-Edit Trace — AI must read file before modifying it
18. Zero-Downtime Migrations — Expand & Contract pattern

Export these rules to your IDE:

ceobe export-rules
# Writes to: .cursor/rules/ceobe.mdc · .windsurf/rules/ceobe.md · .clinerules · .agents/rules/AGENTS.md

Environment Variables

| Variable | Default | Description | |---|---|---| | CEOBE_PLANNER_PROVIDER | — | Provider for Planner brain | | CEOBE_EXECUTOR_PROVIDER | — | Provider for Executor brain | | CEOBE_QA_PROVIDER | (falls back to Planner) | Provider for QA Auditor | | CEOBE_EMBEDDING_PROVIDER | (falls back to Planner) | Provider for RAG embeddings | | CEOBE_*_MODEL | (provider default) | Model override per role | | CEOBE_MAX_BUDGET | 5 | Budget limit in USD (0 = unlimited) | | CEOBE_MAX_TOKENS | 16384 | Max output tokens per call | | CEOBE_SANDBOX | none | docker to isolate execution | | CEOBE_SANDBOX_IMAGE | — | Custom Docker image for sandbox | | CLOUDFLARE_ACCOUNT_ID | — | Cloudflare AI Gateway account | | CLOUDFLARE_GATEWAY_ID | — | Cloudflare AI Gateway ID | | TELEGRAM_BOT_TOKEN | — | Telegram bot token for daemon | | TELEGRAM_ALLOWED_USERS | — | Comma-separated Telegram user IDs |

Runtime Files (.ceobe/)

| File | Purpose | |---|---| | brd.md | Business Requirements Document | | design.md | UI/UX Design Specification | | architecture.md | System Architecture Document | | devops.md | DevOps & Deployment Config | | task.md | Implementation Task Plan | | config.json | Mode config (autonomous/ask) | | ceobe-state.json | Pipeline state & phase tracking | | execution.log | Full execution log | | embeddings.json | Vector index for RAG | | full-text-index.json | BM25 full-text index | | dependency-graph.json | Module dependency map | | plugins/ | Custom plugin directory | | skills/ | Custom project-local skills |

Contributing

Ceobe is designed for extension. The simplest contribution is a new skill:

1. Fork the repository
2. Create skills/<your-skill-name>/ with a .md file
3. Submit a Pull Request

For core contributions, follow the 18 engineering rules enforced by Ceobe itself.

License

MIT · Built by Yoragam1