• 14 MCP tools for hierarchical work item management
• Persistent SQLite database across sessions (Docker volume)
• Role-based workflow: queue → work → review → terminal
• Dependency tracking with blocking, cascade, and unblock detection
• Works with any MCP-compatible client (Claude Code, VS Code, Cursor, etc.)

• Docker installed and running

Reference the Quick Start for detailed installation steps:

See Quick Start Steps 1-3 for Docker image pull, Claude Code registration, and verification.

For non-Claude-Code MCP clients, the server configuration follows the standard MCP format:

{
 "mcpServers": {
 "mcp-task-orchestrator": {
 "command": "docker",
 "args": [
 "run", "--rm", "-i",
 "-v", "mcp-task-data:/app/data",
 "ghcr.io/jpicklyk/task-orchestrator:latest"
 ]
 }
 }
}

The mcp-task-data Docker volume persists the SQLite database across container restarts.

Use manage_items to create individual items, or create_work_tree for hierarchy:

create_work_tree(
 root: { title: "Add user search", summary: "Full-text search for user profiles" },
 children: [
 { title: "Design search API", priority: "high" },
 { title: "Build search index", priority: "medium" },
 { title: "Add search UI", priority: "medium" }
 ],
 dependencyPattern: "linear"
)

This creates a root item with 3 children wired in sequence (each blocks the next).

Items progress via named triggers — never update the role directly:

advance_item(transitions=[{ itemId: "<uuid>", trigger: "start" }])

See Workflow Guide Section 2 for the full trigger table.

Use get_context() with no parameters for a health check of all active work:

get_context() // Returns: active items, blocked items, stalled items

Use get_context(itemId="<uuid>") to inspect a specific item's gate status and expected notes.

get_next_item() // Returns top items ranked by priority then complexity

Use get_next_item(parentId="<uuid>") to scope recommendations to a specific subtree.

1. Create the feature with create_work_tree (root + 3 children, linear dependencies)
2. Start the first child: advance_item(trigger="start") → queue→work
3. Do the work, fill work-phase notes
4. Advance: advance_item(trigger="start") — moves to review (if schema has review-phase notes) or terminal (if not)
5. If in review: fill review notes, then advance_item(trigger="start") → terminal
6. The next child auto-unblocks — check with get_blocked_items(parentId="<root-uuid>")
7. Use get_next_item(parentId="<root-uuid>") to find what's ready
8. When all children complete, the parent auto-cascades to terminal

See API Reference for full tool documentation.

Signal: You find yourself repeating the same workflow instructions to the agent every session — "use advance_item, not raw updates," "check get_context before starting," "fill notes before advancing."

Next: CLAUDE.md-Driven Workflow — embed these conventions so the agent follows them automatically.