AI Coding Tools Orchestrator and Agentic Team Runtime

Five independent systems — an AI Orchestrator, an Agentic Team runtime, an MCP Server, a Context Dashboard, and Graphify (project-to-graph intelligence engine) — that coordinate cloud and local AI coding assistants (Claude, Codex, Gemini, Copilot, Ollama, llama.cpp) to collaborate on software development tasks. Includes enterprise-grade agentic infrastructure with specialized agents, skills library, 34+ MCP tools, project-scoped graph-based context memory, and Graphify's 22-language code analysis with persistent queryable knowledge graphs, interactive visualization, and REST API.

Overview | Architecture | Agentic Infrastructure | System Comparison | Features | Quick Start | Project Structure | Configuration | Deployment | Testing | MCP Server

Overview

AI Coding Tools ships five independent systems in a single repository:

1. The Orchestrator runs step-based workflows where AI agents execute tasks in sequence (implement, review, refine).
2. The Agentic Team runs a free-communication runtime where role-based agents (Project Manager, Architect, Developer, QA, DevOps) discuss a task in turns until the team lead declares the work complete.
3. Graphify turns any project directory into a queryable knowledge graph — classes, functions, imports, call graphs, and design rationale stored in a local SQLite database with FTS5 search.
4. The MCP Server bridges both engines to IDE-based AI assistants.
5. The Context Dashboard visualizes the graph memory. Each system carries its own adapters, configuration, UI, and CLI — they share zero code and zero imports.

Beyond the core engines, we provide a complete Agentic Infrastructure that empowers AI agents:

• 9 Specialized Agents for web, backend, security, DevOps, AI/ML, database, mobile, performance, and documentation
• 22 Reusable Skills across development, testing, security, DevOps, AI/ML, and documentation
• 34+ MCP Tools for code analysis, security scanning, testing, DevOps, and context memory
• Graph Context System with hybrid search (BM25 + semantic) for persistent memory and learning
• Domain Rules encoding best practices for security, database, API design, performance, and AI/ML

[!TIP] Quickstart with the Orchestrator for structured workflows, or the Agentic Team for open-ended collaboration. Both systems benefit from the shared agentic infrastructure and context memory. See QUICKSTART.md for quick setup instructions to get started in ~2 minutes. Or, see #quick-start below for a detailed walkthrough.

Agentic Infrastructure

We provide a rich agentic infrastructure that both the Orchestrator and Agentic Team utilize to empower their AI agents with specialized knowledge, reusable skills, powerful tools, and persistent context memory. This infrastructure is designed to be modular and extensible, allowing for easy addition of new agents, skills, tools, and context nodes as the system evolves.

graph TB
    subgraph "🧠 Agentic Infrastructure"
        direction LR

        subgraph AGENTS["Specialized Agents (9)"]
            WEB[Web Frontend]
            API[Backend API]
            SEC[Security]
            OPS[DevOps]
            ML[AI/ML]
            DB[Database]
        end

        subgraph SKILLS["Skills Library (22)"]
            DEV[Development]
            TEST[Testing]
            SECS[Security]
            DEVOPS[DevOps]
            AIML[AI/ML]
            DOCS[Documentation]
        end

        subgraph TOOLS["MCP Tools (34+)"]
            CODE[Code Analysis]
            SCAN[Security Scan]
            TTOOLS[Testing]
            DTOOLS[DevOps]
            CTX[Context Memory]
        end

        subgraph CONTEXT["Graph Context"]
            GRAPH[(Graph Store)]
            SEARCH[Hybrid Search]
            EMBED[Embeddings]
        end
    end

    AGENTS --> SKILLS
    SKILLS --> TOOLS
    TOOLS --> CONTEXT

[!IMPORTANT] 📚 Click for Full Agentic Infrastructure Documentation

Context System

Both the Orchestrator and Agentic Team maintain independent graph-based context databases for persistent memory and cross-session learning. A unified Context Dashboard aggregates both stores for visualization.

graph TB
    subgraph "Context System Architecture"
        direction TB

        subgraph ORCH_CTX["Orchestrator Context<br/>~/.ai-orchestrator/context.db"]
            OM[models/ — Node & edge schemas]
            OS[store/ — Graph persistence]
            OX[search/ — BM25 + semantic + hybrid]
            OO[ops/ — Analytics, export, pruning, versioning]
        end

        subgraph TEAM_CTX["Agentic Team Context<br/>~/.agentic-team/context.db"]
            TM[models/ — Node & edge schemas]
            TS[store/ — Graph persistence]
            TX[search/ — BM25 + semantic + hybrid]
            TO[ops/ — Analytics, export, pruning, versioning]
        end

        subgraph DASH["Context Dashboard :5003"]
            APP[app.py — Flask aggregator]
            VIZ[templates/ — Interactive visualization]
        end

        subgraph OBS_OUT["Obsidian Export"]
            OBS_V["Obsidian Vault<br/>[[wikilinks]] + graph.json"]
        end
    end

    ORCH_CTX --> DASH
    TEAM_CTX --> DASH
    ORCH_CTX --> OBS_OUT
    TEAM_CTX --> OBS_OUT

    style ORCH_CTX fill:#1a365d,color:#fff
    style TEAM_CTX fill:#1a365d,color:#fff
    style DASH fill:#2d3748,color:#fff
    style OBS_OUT fill:#7C3AED,color:#fff

Node Types — 10 types of knowledge stored in the graph:

Edge Types — 12 relationship types connecting nodes:

Hybrid Search combines three retrieval strategies via Reciprocal Rank Fusion (RRF):

1. BM25 — Keyword-based search using term frequency–inverse document frequency
2. Semantic — Embedding-based similarity using vector cosine distance
3. Hybrid — Fused ranking of BM25 + semantic results using RRF for best-of-both-worlds retrieval

Project-Scoped Context Graphs

Both systems support project-scoped context graphs for full portability. When a user points the system at their project directory, agents automatically scan and build a rich context graph of the codebase.

graph TB
    subgraph "Project Context Scoping"
        direction TB

        USER[User configures PROJECT_PATH] --> SCAN[ProjectScanner]
        SCAN --> PID["project_id = SHA-256[:16] of path"]

        subgraph "Isolated Project Graphs"
            P1["Project A<br/>pid=a1b2c3..."]
            P2["Project B<br/>pid=d4e5f6..."]
            P3["Global Scope<br/>pid='' (no project)"]
        end

        PID --> P1 & P2
        SCAN --> FILES[File Nodes]
        SCAN --> PATTERNS[Pattern Nodes]
        SCAN --> DECISIONS[Decision Nodes]
        SCAN --> EDGES[Relationship Edges]
    end

    style P1 fill:#2b6cb0,stroke:#2c5282,color:#fff
    style P2 fill:#276749,stroke:#22543d,color:#fff
    style P3 fill:#744210,stroke:#975a16,color:#fff

Key features:

• Deterministic IDs: project_id is a SHA-256 prefix of the normalized absolute path — idempotent and reproducible
• Multi-project isolation: Each project gets its own graph partition; queries filter by project_id
• Global scope: Nodes with project_id="" are universal (patterns, reference knowledge) — shared across all projects
• Automatic scanning: ProjectScanner detects languages, frameworks, file structure, and config patterns
• Portability: Set PROJECT_PATH environment variable or settings.project_path in config YAML — the system handles the rest
• Incremental updates: rescan_project() rebuilds the graph atomically; delete_project_graph() cleanly removes all project nodes

[!TIP] Auto-seeding: Run scripts/seed_context_graphs.py to populate both context databases with generic reference knowledge (patterns, mistakes, decisions) on first use. Seed data contains no hallucination-prone fake tasks or conversations — only universally applicable best practices.

[!NOTE] Context Dashboard: Launch with python -m context_dashboard (port 5003) to visualize both context graphs, inspect nodes/edges, and search across all stored knowledge. See context_dashboard/README.md for details.

[!TIP] Obsidian Export: Export any graph as an Obsidian vault for offline interactive exploration. Graphify: graphify export obsidian <path>. Context systems: ContextExporter.export_obsidian(). Each vault includes [[wikilinks]], typed folders, YAML frontmatter, and pre-configured .obsidian/graph.json color groups. See GRAPHIFY.md, ORCHESTRATOR.md, and AGENTIC_TEAM.md.

Skills Library & Agent Definitions

AI coding agents are enhanced with specialized role definitions, reusable skills, and domain rules that are automatically loaded based on context.

graph LR
    subgraph "Agent Ecosystem"
        direction TB

        subgraph CLAUDE[".claude/"]
            CA["agents/ (11)"]
            CS["skills/ (23)"]
            CR["rules/ (11)"]
            CC[CLAUDE.md]
        end

        subgraph CODEX[".codex/"]
            XA["agents/ (13)"]
            XC[config.toml]
            XR[rules/]
        end

        AGENTS_MD[AGENTS.md — Shared instructions]
    end

    CA --> CS
    CA --> CR
    AGENTS_MD --> CLAUDE
    AGENTS_MD --> CODEX

    style CLAUDE fill:#7c3aed,color:#fff
    style CODEX fill:#059669,color:#fff

Claude Agents — 11 specialized agents in .claude/agents/:

Codex Agents — 13 specialized agents in .codex/agents/:

Skills Library — 24 reusable skill templates in .claude/skills/ across 7 categories:

Four additional standalone skills (context-graph-builder, generate-reports, health-check, run-tests) provide operational task automation.

Domain Rules — 11 rule files in .claude/rules/ encoding best practices:

[!NOTE] Agents automatically inherit access to all skills and rules in their scope. When Claude Code is invoked with a specialized agent (e.g., @security-specialist), it loads the agent definition, relevant skills, and applicable domain rules to provide expert-level guidance.

Configuration Files

Architecture

High-Level Overview

graph TD
    subgraph Repository["AI Coding Tools Repository"]
        direction TB

        subgraph Orchestrator["orchestrator/"]
            O_CLI["CLI Shell"]
            O_UI["Web UI<br/>Nuxt 3 + Flask + Socket.IO"]
            O_CORE["Core Engine<br/>Workflow Manager<br/>Task Manager"]
            O_ADAPT["Adapters<br/>Claude | Codex | Gemini<br/>Copilot | Ollama | llama.cpp"]
            O_RESIL["Resilience<br/>Retry | Fallback | Offline"]
            O_OBS["Observability<br/>Prometheus | Logging | Health"]
            O_SEC["Security Module<br/>Validation | Rate Limiting | Audit"]
            O_INFRA["Infra<br/>Cache | Async Executor | Config Manager"]
            O_CONF["orchestrator/config/agents.yaml"]
        end

        subgraph AgenticTeam["agentic_team/"]
            A_CLI["CLI REPL"]
            A_UI["Web UI<br/>Nuxt 3 + Flask + Socket.IO"]
            A_ENGINE["Engine<br/>Free Communication<br/>Lead-Gated Output"]
            A_ADAPT["Adapters<br/>Claude | Codex | Gemini<br/>Copilot | Ollama | llama.cpp"]
            A_FALLBACK["Fallback + Offline"]
            A_CONF["orchestrator/config/agents.yaml"]
        end
    end

    O_CLI --> O_CORE
    O_UI --> O_CORE
    O_CORE --> O_ADAPT
    O_CORE --> O_RESIL
    O_CORE --> O_OBS
    O_CORE --> O_SEC
    O_CORE --> O_INFRA
    O_ADAPT --> ExtCloud["Cloud CLIs<br/>claude | codex | gemini | copilot"]
    O_ADAPT --> ExtLocal["Local Backends<br/>Ollama | llama.cpp"]

    A_CLI --> A_ENGINE
    A_UI --> A_ENGINE
    A_ENGINE --> A_ADAPT
    A_ENGINE --> A_FALLBACK
    A_ADAPT --> ExtCloud
    A_ADAPT --> ExtLocal

    style Orchestrator fill:#1a1a2e,stroke:#16213e,color:#e0e0e0
    style AgenticTeam fill:#1a2e1a,stroke:#162e16,color:#e0e0e0

Orchestrator Workflow Execution

The Orchestrator processes tasks through a configurable pipeline of AI agents. Each step in a workflow maps to a specific agent and role.

sequenceDiagram
    participant User
    participant CLI as CLI / Web UI
    participant Engine as Core Engine
    participant WF as Workflow Manager
    participant Codex as Codex Adapter
    participant Gemini as Gemini Adapter
    participant Claude as Claude Adapter
    participant FB as Fallback Manager

    User->>CLI: Submit task
    CLI->>Engine: execute(task, workflow="default")
    Engine->>WF: load workflow steps

    WF->>Codex: Step 1 -- implement
    alt Codex unavailable
        Codex-->>FB: error
        FB->>FB: route to local-code
    end
    Codex-->>WF: implementation

    WF->>Gemini: Step 2 -- review
    Gemini-->>WF: review feedback

    WF->>Claude: Step 3 -- refine
    Claude-->>WF: refined code

    WF-->>Engine: final result
    Engine-->>CLI: display output
    CLI-->>User: code + report

Agentic Team Communication Flow

The Agentic Team uses free role-to-role communication. Agents speak in turns, address each other by role, and the team lead decides when the task is complete.

sequenceDiagram
    participant User
    participant PM as Project Manager (Lead)
    participant Arch as Software Architect
    participant Dev as Software Developer
    participant QA as QA Engineer
    participant DevOps as DevOps Engineer

    User->>PM: "Build a REST API with auth"
    PM->>Arch: Define architecture and constraints
    Arch->>Dev: Provide interface specs
    Dev->>Dev: Implement code
    Dev->>QA: Request quality review
    QA->>Dev: Report edge cases
    Dev->>Dev: Fix issues
    Dev->>DevOps: Request deployment review
    DevOps->>PM: Confirm operational readiness
    PM->>User: Final consolidated response

Adapter Resolution Flow

Both systems resolve which AI backend to use at runtime. The adapter layer abstracts cloud CLIs and local model servers behind a common interface.

flowchart TD
    REQ[Incoming Task Step] --> CHECK{Agent Enabled?}
    CHECK -->|Yes| HEALTH{Health Check}
    CHECK -->|No| SKIP[Skip Agent]

    HEALTH -->|Healthy| EXEC[Execute via Adapter]
    HEALTH -->|Unhealthy| FB{Fallback Configured?}

    FB -->|Yes| LOCAL[Route to Local Adapter<br/>Ollama / llama.cpp]
    FB -->|No| ERR[Raise AgentUnavailableError]

    EXEC --> PARSE[Parse CLI Output]
    LOCAL --> PARSE
    PARSE --> RESULT[Return AgentResponse]

    style EXEC fill:#2b6cb0,stroke:#2c5282,color:#fff
    style LOCAL fill:#276749,stroke:#22543d,color:#fff
    style ERR fill:#9b2c2c,stroke:#742a2a,color:#fff

Local model execution semantics and limitation

Local models are deeply integrated for routing, offline mode, and fallback, but they are not direct workspace editors in the current implementation.

Best use for local models: offline drafting, review feedback, and cloud-to-local fallback continuity.

[!CAUTION] While it is possible to make local LLMs directly edit files (e.g., via a file-editor tool), this approach is currently disabled to prevent unintended destructive changes. Local adapters are advisory — they provide text output that the Orchestrator can use to inform the next steps, but they do not have direct write access to the workspace. This design choice prioritizes safety and predictability while still leveraging local models for their strengths in drafting and feedback. The hard part is not feasibility, it’s safety and reliability: permissions, diff constraints, validation/tests before write, rollback, and preventing bad edits.

Technology Stack Overview

graph LR
    subgraph Backend
        PY[Python 3.8+]
        FL[Flask 3.x]
        SIO[Socket.IO 4.x]
        PD[Pydantic 2.x]
        CL[Click 8.x]
    end

    subgraph Frontend
        VUE[Vue 3]
        NUXT[Nuxt 3]
        TW[Tailwind CSS 3.x]
        MON[Monaco Editor]
        PIN[Pinia]
    end

    subgraph Observability
        PROM[Prometheus]
        GRAF[Grafana]
        SL[structlog]
    end

    subgraph Infrastructure
        DOCK[Docker]
        K8S[Kubernetes]
        TF[Terraform]
    end

    PY --> FL --> SIO
    PY --> PD
    PY --> CL
    VUE --> NUXT --> TW
    VUE --> MON
    VUE --> PIN
    PROM --> GRAF
    DOCK --> K8S

    style Backend fill:#1a1a2e,stroke:#16213e,color:#e0e0e0
    style Frontend fill:#1a2e1a,stroke:#162e16,color:#e0e0e0
    style Observability fill:#2e1a1a,stroke:#2e1616,color:#e0e0e0
    style Infrastructure fill:#1a1a2e,stroke:#16213e,color:#e0e0e0

System Comparison

The two systems serve different collaboration models. Choose based on your use case.

Feature Highlights

Orchestrator (orchestrator/)

Agentic Team (agentic_team/)

Graphify (graphify/)

Quick Start

Prerequisites

• Operating System: Linux, macOS, or Windows (WSL recommended)
• Python: 3.8 or higher
• Node.js: 20+ (for Web UI)
• Memory: Minimum 4GB RAM
• Disk Space: 1GB for installation + workspace
• Network: Required for AI CLI tools and updates
• Claude Code: Installed, setup, and signed in on your machine (Required for any workflows using Claude Code - if you run claude in terminal and it works, you're good)
• OpenAI Codex: Installed and authenticated (if using Codex agent, try running codex and see if it responds)
• Google Gemini CLI: Installed and authenticated (if using Gemini agent, try gemini --version to verify)
• GitHub Copilot CLI: Installed and authenticated (if using Copilot agent, try copilot --version to verify)
• Llama.cpp or Ollama: If using local LLM agents, ensure they are installed and configured properly (try running ollama list or llamacpp --help to verify)
• Optional: Docker and Docker Compose for containerized setup

Install

git clone https://github.com/hoangsonww/AI-Agents-Orchestrator.git
cd AI-Agents-Orchestrator

python3 -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate

pip install -r requirements.txt
chmod +x ai-orchestrator

Run the Orchestrator

# Interactive shell
./ai-orchestrator shell

# One-shot task
./ai-orchestrator run "Create a Python REST API" --workflow default

# Start the Web UI
make run-ui
# Then open http://localhost:5001

Run the Agentic Team

# Interactive REPL
./ai-orchestrator agentic-shell

# With options
./ai-orchestrator agentic-shell --max-turns 16 --offline

# Start the Web UI
make run-agentic-ui
# Then open http://localhost:5002

Verify Installation

./ai-orchestrator --help        # Show all commands
./ai-orchestrator agents        # List available agents
./ai-orchestrator workflows     # List available workflows
./ai-orchestrator validate      # Validate configuration

Project Structure

AI-Coding-Tools/
|
|-- .claude/                         # Claude Code agentic infrastructure
|   |-- CLAUDE.md                    # Main Claude instructions (imports AGENTS.md)
|   |-- settings.json                # Project settings and permissions
|   |-- agents/                      # 11 specialized agent definitions
|   |   |-- web-frontend.md
|   |   |-- backend-api.md
|   |   |-- security-specialist.md
|   |   |-- devops-infrastructure.md
|   |   |-- ai-ml-engineer.md
|   |   |-- database-architect.md
|   |   |-- mobile-developer.md
|   |   |-- performance-engineer.md
|   |   |-- documentation-writer.md
|   |   |-- code-reviewer.md
|   |   +-- test-runner.md
|   |-- skills/                      # 23 reusable skill templates
|   |   |-- development/             # 6 skills (react, REST, async, DB, GraphQL, errors)
|   |   |-- testing/                 # 4 skills (unit, integration, TDD, perf)
|   |   |-- security/                # 4 skills (validation, auth, secure-coding, vuln)
|   |   |-- devops/                  # 3 skills (Docker, CI/CD, K8s)
|   |   |-- ai-ml/                   # 3 skills (embeddings, LLM, RAG)
|   |   |-- documentation/           # 3 skills (API docs, arch docs, code docs)
|   |   |-- generate-reports/        # Standalone: report generation
|   |   |-- health-check/            # Standalone: system health checks
|   |   |-- run-tests/               # Standalone: test suite execution
|   |   +-- context-graph-builder/   # Standalone: context graph operations
|   +-- rules/                       # 11 domain rule files
|       |-- adapters.md
|       |-- api-design.md
|       |-- testing.md
|       |-- performance.md
|       |-- config.md
|       |-- ai-ml.md
|       |-- observability.md
|       |-- frontend.md
|       |-- ci-cd.md
|       |-- security.md
|       +-- database.md
|
|-- .codex/                          # Codex agentic infrastructure
|   |-- config.toml                  # Codex project configuration
|   |-- agents/                      # 13 specialized agent definitions (.toml)
|   |   |-- code-reviewer.toml
|   |   |-- explorer.toml
|   |   |-- security-specialist.toml
|   |   |-- web-frontend.toml
|   |   |-- devops-infrastructure.toml
|   |   |-- implementer.toml
|   |   |-- database-architect.toml
|   |   |-- performance-engineer.toml
|   |   |-- test-runner.toml
|   |   |-- ai-ml-engineer.toml
|   |   |-- backend-api.toml
|   |   |-- documentation-writer.toml
|   |   +-- mobile-developer.toml
|   |-- hooks/                       # Git hook integrations
|   +-- rules/                       # Codex-specific rules
|
|-- mcp_server/                      # MCP server (FastMCP 3.x) — 34+ tools
|   |-- server.py                    # Server entry point + core tools
|   |-- engines.py                   # Engine adapters for both systems
|   |-- repl.py                      # Interactive REPL mode
|   |-- tools/                       # Tool modules by category
|   |   |-- orchestrator_tools.py    # Orchestrator execution tools
|   |   |-- agentic_team_tools.py    # Agentic team execution tools
|   |   |-- shared_tools.py          # Shared utility tools
|   |   |-- code_analysis.py         # 4 code analysis tools
|   |   |-- security_tools.py        # 4 security scanning tools
|   |   |-- testing_tools.py         # 4 testing tools
|   |   |-- devops_tools.py          # 5 DevOps tools
|   |   +-- context_tools.py         # 7 context memory tools
|   +-- resources/                   # MCP resource definitions
|
|-- orchestrator/                    # Self-contained orchestrator system
|   |-- __init__.py
|   |-- adapters/                    # AI agent adapters
|   |   |-- base.py                  # Abstract base adapter
|   |   |-- claude_adapter.py        # Claude Code CLI
|   |   |-- codex_adapter.py         # OpenAI Codex CLI
|   |   |-- gemini_adapter.py        # Google Gemini CLI
|   |   |-- copilot_adapter.py       # GitHub Copilot CLI
|   |   |-- ollama_adapter.py        # Ollama local models
|   |   |-- llama_cpp_adapter.py     # llama.cpp / OpenAI-compatible
|   |   +-- cli_communicator.py      # Robust CLI subprocess handling
|   |