hugoduncan/mcp-tasks

mcp-tasks

Task-based workflow management for AI agents via Model Context Protocol (MCP).

What is mcp-tasks?

mcp-tasks enables you to manage development tasks in EDNL files and have AI agents execute them. Unlike todo tools, mcp-tasks integrates task planning with execution—agents don't just track tasks, they complete them.

Key Benefits:

• Persistent Planning: Tasks survive across chat sessions in EDNL files
• Category-Based Organization: Group tasks by type (features, bugfixes, refactoring) with custom execution strategies
• Story-Based Development: Break down epics into tasks with dedicated workflows
• Automatic Branch Creation: Tasks automatically get isolated branches for clean development
• Git Worktree Management: Parallel development across multiple tasks without branch switching
• Audit Trail: Completed tasks automatically archived with full context
• Flexible Workflows: Supports both git-tracked and standalone task management

For a deeper understanding, see .

Quick Start

Installation

Install native binaries (recommended):

# Unix (Linux/macOS)
curl -sL https://raw.githubusercontent.com/hugoduncan/mcp-tasks/master/install | bash

This installs:

• mcp-tasks CLI for terminal task management
• mcp-tasks-server MCP server for AI agent integration

Note: Windows native binaries are not currently supported due to file locking complexities. Windows users can use the Clojure git dependency installation method (see ).

See for alternative installation methods (Clojure git dependency, building from source).

Install Claude Code Skill (Optional):

claude skill add plugins/mcp-tasks-skill

The skill provides:

• Comprehensive guidance on available tools and prompts
• Task lifecycle walkthroughs and workflow patterns
• Git integration best practices
• Error recovery and troubleshooting

Alternative for non-Claude agents: Add @plugins/mcp-tasks-skill/skills/story-and-tasks/SKILL.md to your agent's context file for similar guidance.

Configuration

Configure Claude Code:

claude mcp add mcp-tasks -- /usr/local/bin/mcp-tasks-server

Initialize task directories in your project:

mkdir -p .mcp-tasks/category-prompts

# Optional: Initialize as git repository for version control
cd .mcp-tasks && git init && git commit --allow-empty -m "Initialize task tracking" && cd ..

Your First Task

Create a task:

Ask your AI agent:

Create a simple category task to add README badges for build status

The agent will use the add-task tool and return a task-id (e.g., 123).

Execute the task:

/mcp-tasks:execute-task 123

The agent will:

• Validate the task and set up the execution environment
• Follow the "simple" category workflow: analyze, implement, commit
• Mark the task as complete

Review:

# Check what was completed
git log -1 --stat

# View completed tasks using select-tasks tool in Claude Code

Core Concepts

Tasks: A unit of work an agent can complete without exceeding its context limits. Stored as EDN records with fields: :id, :category, :title, :description, :status, :type, :design, :meta, :relations.

Categories: Task organization by type/workflow. Each category has a prompt file (e.g., .mcp-tasks/prompts/simple.md) defining execution steps. Built-in categories include simple, medium, large.

Stories: Large features broken down into multiple related tasks. Story workflows include refinement, task creation, and sequential execution.

Prompts: MCP prompts that guide agents through task execution (e.g., /mcp-tasks:category-simple). Each category has a corresponding prompt.

See and for complete details.

Common Workflows

Task-ID Based Workflow

# Ask agent to create tasks
"Create a simple task to add user profile endpoint"
# Agent uses add-task tool, returns task-id: 123

"Create a medium task to fix memory leak in worker process"
# Agent uses add-task tool, returns task-id: 124

# Execute specific tasks by ID (direct control over execution order)
/mcp-tasks:execute-task 124     # Agent fixes bug, commits
/mcp-tasks:execute-task 123     # Agent adds endpoint, commits

# Check completion history in .mcp-tasks/complete.ednl

Category-Based Workflow

# Add tasks using add-task tool in Claude Code
# category: "feature", title: "Add user profile endpoint"
# category: "bugfix", title: "Fix memory leak in worker process"

# Execute by category (sequential processing from queue)
/mcp-tasks:category-bugfix      # Agent fixes bug, commits
/mcp-tasks:category-feature     # Agent adds endpoint, commits

# Check completion history in .mcp-tasks/complete.ednl

Story-Based Workflow

# Create a story using add-task tool
# category: "large", type: "story"
# title: "User Authentication System"
# description: "Implement JWT-based authentication..."

# Refine the story interactively
/mcp-tasks:refine-task "User Authentication System"

# Break down into tasks
/mcp-tasks:create-story-tasks "User Authentication System"

# Execute tasks sequentially
/mcp-tasks:execute-story-child "User Authentication System"

See for advanced patterns including git worktrees and parallel task execution.

Documentation

• - Deep dive into how mcp-tasks helps you work with AI agents
• - Complete installation instructions for all platforms
• - Task execution workflows, story workflows, and advanced patterns
• - Config file options, git integration, worktree management
• - Custom categories, prompt templates, task metadata, MCP resources
• - CLI commands, output formats, and usage
• - Development setup, architecture, testing, and building
• - Terms and concepts
• - Build instructions for native binaries and JAR

Status & License

Status: Alpha. Core functionality stable; API may evolve.

License: EPL-2.0 (see LICENSE file)

Issues & Feedback: https://github.com/hugoduncan/mcp-tasks/issues

Changelog: