MCP Task Orchestrator

Stop losing context. Start building faster.

An orchestration framework for AI coding assistants that solves context pollution and token exhaustion - enabling your AI to work on complex projects without running out of memory.

---

The Problem

AI assistants suffer from context pollution - a well-documented challenge where model accuracy degrades as token count increases. This "context rot" stems from transformer architecture's quadratic attention mechanism, where each token must maintain pairwise relationships with all others.

The Impact: As your AI works on complex features, it accumulates conversation history, tool outputs, and code examples. By task 10-15, the context window fills with 200k+ tokens. The model loses focus, forgets earlier decisions, and eventually fails. You're forced to restart sessions and spend 30-60 minutes rebuilding context just to continue.

Industry Validation: Anthropic's research on context management confirms production AI agents "exhaust their effective context windows" on long-running tasks, requiring active intervention to prevent failure.

Traditional approaches treat context windows like unlimited memory. Task Orchestrator recognizes they're a finite resource that must be managed proactively.

The Solution

Task Orchestrator implements industry-recommended patterns from Anthropic's context engineering research: persistent external memory, summary-based context passing, and sub-agent architectures with clean contexts.

How it works:

• Persistent memory (SQLite) stores project state outside context windows
• Summary-based passing - Tasks create 300-500 token summaries instead of passing 5-10k full contexts
• Sub-agent isolation - Specialists work with clean contexts, return condensed results
• Just-in-time loading - Fetch only what's needed for current work

Result: Scale to 50+ tasks without hitting context limits. Up to 90% token reduction (matching Anthropic's 84% benchmark). Zero time wasted rebuilding context.

---

Key Features

• ✅ Persistent Memory - AI remembers project state, completed work, and decisions across sessions
• ✅ Token Efficiency - Up to 90% reduction via summary-based context passing
• ✅ Hierarchical Tasks - Projects → Features → Tasks with dependency tracking
• ✅ Template System - 9 built-in workflow templates with decision frameworks and quality gates
• ✅ Event-Driven Workflows - Automatic status progression based on your config
• ✅ Sub-Agent Orchestration - Specialist routing for complex work (Claude Code)
• ✅ Skills & Hooks - Lightweight coordination and workflow automation (Claude Code)
• ✅ MCP Protocol Support - Core persistence and task management work with any MCP client

📖 Deep dive: See Agent Architecture Guide for token efficiency comparison and Developer Architecture for technical details.

---

Quick Start

Option A: Plugin Installation (Recommended for Claude Code)

Easiest way - Install everything (MCP server, skills, subagents, hooks) in one step:

1. Clone this repository:

   git clone https://github.com/jpicklyk/task-orchestrator.git
   cd task-orchestrator
   

2. Add the local marketplace:

   /plugin marketplace add ./
   

3. Install the plugin:

   /plugin install task-orchestrator@task-orchestrator-marketplace
   

4. Restart Claude Code

5. Initialize your project:

   setup_project
   

Note: Once this repository is published on GitHub, you'll be able to use:

/plugin marketplace add jpicklyk/task-orchestrator
/plugin install task-orchestrator

See Plugin Installation Guide for detailed instructions and troubleshooting.

Option B: Manual MCP Installation

For other MCP clients or custom setup:

1. Install via Docker:

   docker pull ghcr.io/jpicklyk/task-orchestrator:latest
   

2. Configure your AI platform:

   Claude Code:

   claude mcp add-json task-orchestrator '{"type":"stdio","command":"docker","args":["run","--rm","-i","-v","mcp-task-data:/app/data","-v",".:/project","-e","AGENT_CONFIG_DIR=/project","ghcr.io/jpicklyk/task-orchestrator:latest"]}'
   

   This single command works across all platforms (macOS, Linux, Windows).

   Other MCP clients: Task Orchestrator's core MC

...