AgentRQ MCP

Overview

AgentRQ is an open-source task manager MCP server that gives AI agents a shared, persistent task queue — a "GitHub Issues for agents" that survives across sessions and clients. The project was featured on Show HN with the tagline "Task Manager for AI Agents (MCP, Opensource)".

Most agent frameworks have session-local memory: tasks disappear when the session ends. AgentRQ solves this by exposing durable task primitives — create, assign, prioritize, complete — through the Model Context Protocol, so any MCP client (Claude Code, Cursor, Zed) can participate in a shared work queue.

Features

• Persistent task storage: tasks survive session restarts
• Multi-client: any MCP client can create, read, and update tasks
• Priorities and labels: first-class task prioritization
• Assignment: tasks can be assigned to specific agent identities
• Comment threads: agents can leave notes on tasks
• Local-first: SQLite by default, Postgres available

Installation

# Install via npx
npx -y @agentrq/mcp

# Or install globally
npm install -g @agentrq/mcp

Configuration

Add to your MCP client configuration:

{
  "mcpServers": {
    "agentrq": {
      "command": "npx",
      "args": ["-y", "@agentrq/mcp"],
      "env": {
        "AGENTRQ_DB_PATH": "~/.agentrq/tasks.db",
        "AGENTRQ_DEFAULT_ASSIGNEE": "main-agent"
      }
    }
  }
}

Available Tools

Tool	Description
create_task	Create a new task with title, description, and priority
list_tasks	List tasks with filters (status, assignee, label)
get_task	Get full details of a specific task
update_task	Update title, description, status, or priority
assign_task	Assign a task to an agent identity
complete_task	Mark a task as complete with optional result
add_comment	Add a comment to a task's thread
search_tasks	Full-text search across task titles and descriptions

Example Usage

Create a task from Claude Code

Ask Claude: "Create a task: Refactor the auth module. Priority: high. Assign to me."

Under the hood, Claude calls:

{
  "tool": "create_task",
  "args": {
    "title": "Refactor the auth module",
    "description": "Consolidate JWT and session-based auth into a single middleware.",
    "priority": "high",
    "assignee": "claude-code",
    "labels": ["refactor", "auth"]
  }
}

Cross-session coordination

A morning session creates planning tasks. An afternoon session picks them up:

# Pseudocode — what AgentRQ exposes as MCP tools
tasks = mcp.list_tasks(status="pending", assignee="claude-code")
for task in tasks:
    result = do_work(task)
    mcp.complete_task(task.id, result=result.summary)

Setup for Cursor

1. Open Cursor Settings → Tools & MCP
2. Add the AgentRQ configuration from above
3. Restart Cursor
4. Verify tools are available in the MCP panel

Pros

• ✅ Durable across sessions: tasks persist when the conversation ends
• ✅ Multi-client: Claude Code, Cursor, Zed, and others can share the queue
• ✅ Simple SQLite backend: no external service needed for small teams
• ✅ Postgres option: scales up when needed
• ✅ First-class comments: agents can leave notes for other agents
• ✅ Local-first: full functionality without cloud

Cons

• ❌ Small community: ~40 GitHub stars, minimal contributor base
• ❌ No real-time sync: tasks update on write, not push
• ❌ No auth: multi-user setups require external protection
• ❌ Task schema is opinionated: not customizable without source changes
• ❌ No web UI: all access is through MCP tools

When to Use

Use AgentRQ MCP when:

• You run long-running multi-session agents and need a shared task state
• Multiple agents need to coordinate without a full orchestration framework
• You want lightweight task tracking for agent-driven workflows
• You're building on top of MCP and want tasks as a primitive

Prefer alternatives when:

• You need real-time coordination — use a proper orchestration framework like LangGraph
• You need auth and multi-tenancy — use a hosted task system (Linear, Jira)
• You only need session-local memory — use Zep or Mem0

Resources

• GitHub Repository
• Show HN discussion