OpenClaw Task Queue v2.0

Persistent task queue for AI agents. 5 tools. Survives restarts. Multi-agent assignment. Priority scheduling.

What's New in v2.0

• TaskFlow Engine -- Priority queue backed by Cloudflare D1 (SQLite at the edge). Tasks persist across agent restarts, session crashes, and PC reboots.
• PriorityLane -- Four priority lanes (critical/high/normal/low) with configurable lane weights. Critical tasks always jump the queue.
• DependencyGraph -- Link tasks with depends_on to create execution chains. A task only becomes available when all its dependencies are complete.
• TimeoutGuard -- Tasks assigned to an agent but not completed within a configurable timeout are automatically returned to the queue for reassignment.

Quick Start

{
  "openclaw-task-queue": {
    "type": "streamable-http",
    "url": "https://openclaw-task-queue-mcp.yagami8095.workers.dev/mcp"
  }
}

Add to Claude Desktop, Cursor, Windsurf, VS Code, or any MCP-compatible client. Zero install. Works immediately.

Detection Triggers

This skill activates when a user says any of these:

• "add this to the queue", "create a task", "remember to do", "queue this for later"
• "what's in the queue", "pending tasks", "show task list", "what needs to be done"
• "assign this task", "give to agent X", "delegate to worker", "pull next task"
• "mark as done", "complete task", "task finished", "close this task"
• "task stats", "queue statistics", "completion rate", "how many tasks done", "backlog count"
• "task depends on", "chain tasks", "run after X is done", "task dependency"

Named Protocols

TaskFlow Engine

The core queue engine built on Cloudflare D1 for persistent, edge-native task management:

Task lifecycle:

created → queued → assigned → in_progress → completed
                ↓                ↓
             cancelled        failed → retry → queued

Task schema:

PriorityLane Rules

How tasks are ordered when an agent calls assign_task:

Within the same priority, tasks are ordered FIFO (oldest first). An agent requesting the "next task" always gets the highest-priority available unassigned task.

DependencyGraph Rules

Task chaining with depends_on:

1. Task B with depends_on: [task_A_id] stays in queued status until task A reaches completed
2. If task A fails or is cancelled, task B is set to blocked with reason
3. Circular dependencies are rejected at create_task time with an error
4. A task can have multiple dependencies: all must complete before it unlocks
5. Diamond dependencies (A→B, A→C, B→D, C→D) are supported

TimeoutGuard Settings

Automatic reassignment of stalled tasks:

Timeout starts when task enters in_progress. An agent must call complete_task before timeout or the task is reassigned.

Tools (5)

create_task -- Create Task via TaskFlow Engine

Create a new task with all required fields. Returns task ID for subsequent assign/complete calls.

Wrong / Right:

WRONG: Storing tasks in an AI agent's context window (lost on session end)
RIGHT: create_task({ title: "Analyze Q1 revenue", priority: "high",
                     description: "Pull Stripe data, compare to Q4, write 1-page summary",
                     tags: ["revenue", "reporting"], agent_id: "agent-7f3a" })
       -> Task stored in D1. Survives restarts. Agent 7f3a gets it on next poll.
       -> Returns: { id: "task-9b2f", status: "assigned", created_at: "..." }

WRONG: Creating tasks without descriptions (agents can't act on a title alone)
RIGHT: Always include a detailed description with context, expected output, and any constraints.
       The description is what the agent reads to understand what to do.

get_tasks -- Task List with Filters

List tasks filtered by status, agent, priority, or tag. Returns paginated results.

Wrong / Right:

WRONG: Asking an agent "what are your tasks" when there's no persistent task list
RIGHT: get_tasks({ status: "queued", priority: "high" })
       -> [{ id: "task-9b2f", title: "Analyze Q1 revenue", priority: "high",
              created_at: "2h ago", depends_on: [] },
           { id: "task-4c1a", title: "Draft investor update", ... }]
       -> Filtered, sorted by priority. Ready for agent to pick up.

WRONG: Fetching all tasks with no filters and overwhelming the context window
RIGHT: Always filter by status. get_tasks({ status: "queued" }) for workable tasks.
       Use get_task_stats for aggregate counts.

assign_task -- Task Assignment

Assign a specific task to an agent, or pull the highest-priority unassigned task for an agent. Sets status to in_progress.

Parameters: task_id (optional — if omitted, pulls next by priority), agent_id, role_hint (optional)

complete_task -- Task Completion

Mark a task as complete with an output summary. Sets status to completed, records output, unlocks dependent tasks.

Parameters: task_id, output (text summary of what was done and the result), duration_sec (optional)

get_task_stats -- Queue Analytics

Get aggregate statistics about the task queue: total tasks by status, throughput, completion rate, average duration, and backlog depth.

Returns: counts by status, tasks_completed_today, tasks_created_today, avg_duration_min, completion_rate_pct, backlog_by_priority

Security & Privacy

• Queue isolation -- Your tasks are scoped to your Pro key. No agent or user can access another user's queue.
• Task content not indexed -- Task descriptions and metadata are stored in D1 but never indexed for search or analytics by OpenClaw.
• Soft delete -- cancelled and completed tasks are retained for 90 days for audit purposes, then purged. There is no permanent deletion before 90 days.
• Agent ID is user-defined -- We do not validate agent IDs. You can use any string identifier. No authentication is required between agents and the queue (access control is via your Pro key).
• Edge isolation -- Each API call runs in a Cloudflare Workers V8 isolate. No shared state between users.
• HTTPS only -- All connections are TLS 1.3. HTTP is rejected.

Pricing

Get Pro Key: https://buy.stripe.com/4gw5na5U19SP9TW288

The OpenClaw Intelligence Stack

All 9 servers share one Pro key. $9/mo = 49 tools.