initiative

Initiative

Start it before bed. Wake up to a codebase that's been analyzed by 8 specialist agents, tasks prioritized and executed in parallel worktrees, with commits ready for review.

Initiative is an autonomous task orchestration plugin for Claude Code. It turns Claude into a never-stopping background worker that continuously pulls tasks from a priority queue, executes them in isolated git worktrees, and when the queue runs dry, convenes a panel of 8 expert agents to analyze your project and generate new high-impact tasks automatically. It doesn't just fix problems -- it builds forward.

How It Works

Initiative runs as an MCP server that provides task management tools to Claude Code. The orchestration loop is simple:

1. Check for tasks -- Pull the highest-priority pending task (respecting dependency order)
2. Execute in parallel -- Dispatch up to 3 subagents in isolated git worktrees for concurrent execution
3. Analyze -- When the queue is empty, convene an 8-agent expert panel to explore the project and generate 3-5 new tasks
4. Repeat -- Never stop unless the user says so

Installation

# Clone the repository
git clone https://github.com/ianmcvann/initiative.git
cd initiative

# Add as a Claude Code plugin
claude plugin add ./

Marketplace (Coming Soon)

# Add the Initiative marketplace
claude plugin marketplace add ianmcvann/initiative

# Install Initiative plugin
claude plugin install initiative@initiative

Start working!

/initiative:start

Usage

Initiative provides seven commands:

/initiative:start [direction]

Start autonomous orchestration. Optionally provide a direction to seed the initial task:

/initiative:start
/initiative:start "Add authentication to the API"

/initiative:stop

Stop the orchestration loop and show a final status summary:

/initiative:stop

/initiative:status

Check the current state of the task queue:

/initiative:status

Shows total tasks, counts by status, average completion time, throughput, and titles of any in-progress tasks.

/initiative:add [task description]

Add a task to the queue without starting the orchestration loop:

/initiative:add "Refactor the database layer to use connection pooling"

/initiative:clear

Clear all pending tasks from the queue:

/initiative:clear

/initiative:view [task_id]

View details of a specific task by ID:

/initiative:view 42

/initiative:history

Show completed and failed task history:

/initiative:history

Expert Panel

When the task queue is empty, Initiative doesn't just sit idle -- it convenes a panel of 8 specialist agents that analyze your project in parallel, each from a different perspective. This is what makes Initiative more than a task runner: it's a team of experts that continuously discovers what your project needs next.

Expert	Focus
Security	Vulnerabilities, input validation gaps, exposed secrets, insecure defaults
Architecture	Code structure, coupling, separation of concerns, scalability bottlenecks
Testing	Coverage gaps, missing edge cases, flaky tests, untested error paths
Documentation	Missing or outdated docs, undocumented APIs, incomplete guides
Performance	N+1 queries, unnecessary computation, missing caching, memory leaks
Product Management	Missing features, incomplete workflows, user value gaps, feature prioritization
UX	Confusing interfaces, poor feedback, accessibility, inconsistent patterns
Marketing	Value proposition, examples, comparisons with alternatives, distribution

After all 8 experts report back, Initiative synthesizes their findings: deduplicates overlapping concerns, prioritizes across domains (critical security > product gaps > architecture > UX > testing > docs > performance > marketing), chains related tasks with dependencies, and generates 3-5 concrete, actionable tasks. At least half of generated tasks are additive -- new features and capabilities, not just fixes.

Architecture

src/initiative/
  models.py      # Task dataclass and TaskStatus enum
  database.py    # SQLite task store with priority queue
  server.py      # FastMCP server exposing task management tools

Task Store -- SQLite database with WAL journaling. Tasks have a title, description, priority (0-1000, higher = more urgent), status, tags, dependency tracking, auto-retry support, and timestamps. The priority queue returns the highest-priority pending task first (ties broken by creation time), skipping any task whose dependencies haven't completed yet.

MCP Server -- Built on FastMCP, exposes 16 tools over stdio transport: