ai-orchestrator

README · Architecture · Agents · Skills & Commands · Plugins

Portable AI developer setup: Claude plans, local Ollama executes.

Works with any project: TypeScript, Python, Flutter, Swift, C++. All orchestration is pure Bash and jq.

/implement triggers a multi-layer smart pipeline:

Layer 0  TRIAGE      → detects domain, chooses route
Layer 1  PLAN        → planner + pre-review (approach approval)
Layer 2  CODE        → coder (Ollama) + build check
Layer 3  GATE        → fast review → deep review (if needed)
Layer 4  FIX LOOP    → error-coordinator, max 3 rounds
Layer 5  FINALIZE    → token savings tracked

Triage auto-detects the task domain (api, docker, security, database, testing, etc.) and loads the right plugins, agents, and standards — no manual configuration needed.

For focused tasks (optimize Dockerfile, run security audit, generate tests) triage routes directly to the matching plugin, skipping the planner and saving tokens. For composite tasks it runs the full pipeline with domain expertise pre-loaded.

Claude orchestrates and plans. A local Ollama model writes and reviews the code. Details: Architecture · Agents

/implement Add JWT authentication to the REST API
/implement Optimize the Dockerfile with multi-stage build
/implement Refactor the user service to reduce complexity
/implement Fix this crash: TypeError cannot read property of undefined
/implement Generate tests for the payment module

The pipeline detects the domain, loads the right expertise, and runs automatically.

• Claude Code CLI
• Ollama installed and running
• jq (install.sh installs it automatically)

curl -sSL https://raw.githubusercontent.com/Mybono/ai-orchestrator/main/scripts/install.sh | bash

git clone https://github.com/Mybono/ai-orchestrator ~/Projects/ai-orchestrator
cd ~/Projects/ai-orchestrator
./scripts/install.sh

Both methods run scripts/install.sh automatically to configure your local system (creating symlinks in ~/.claude/ and configuring Ollama models for your hardware).

Model routing is controlled by llm-config.json in the repo root:

{
  "models": {
    "coder":        "hf.co/bartowski/Qwen2.5-Coder-14B-Instruct-GGUF:IQ4_XS",
    "reviewer":     "qwen2.5-coder:7b",
    "pre-reviewer": "qwen2.5-coder:7b",
    "debugger":     "qwen2.5-coder:7b",
    "devops":       "qwen2.5-coder:7b",
    "quick-coder":  "qwen2.5-coder:7b",
    "commit":       "qwen2.5-coder:7b",
    "triage":       "llama3.1:8b",
    "embedding":    "mxbai-embed-large"
  }
}

Changing a model name takes effect immediately without restarting anything. See Architecture → Model Configuration for details.

All commands and agents: Skills & Commands · Agents · Plugins

Domain-specific extensions that add slash commands to the orchestrator. Each plugin in plugins/ handles a specific area — accessibility, database work, Docker, Kubernetes, testing, security, and more.

Full list with trigger keywords and paired agents: Plugins

install.sh adds shell aliases for these commands automatically:

local-commit              # stage all changes, generate a commit message via Ollama, confirm and commit
open-pr                   # generate a PR title and description via Ollama, optionally create it via gh
stats [day|week|month]    # show token savings summary

The orchestrator tracks every Ollama call. View estimated savings vs Claude Sonnet pricing:

/stats week

───────────────────────────────
 ai-orchestrator savings
 Period: this week
 Runs: 12
 Tokens saved: ~186k
 Estimated saving: $7.20
 ───────────────────────────────

To apply orchestration rules in any project:

cp ~/.claude/ai_rules.md ~/Projects/your-project/ai_rules.md

Compatible with .cursorrules and .clauderules.

cd ~/Projects/ai-orchestrator && git pull

Changes apply immediately via symlinks, so you do not need to reinstall.

README · Architecture · Agents · Skills & Commands · Plugins# Testing shellcheck