AgentRQ is a self-hosted task manager shared between people and AI agents. A workspace holds a queue of tasks, and an agent connected to that workspace over the Model Context Protocol (MCP) pulls its assigned tasks, updates their status, replies in the task thread, asks a person for a decision, and fires events when work completes. The person follows along from a web dashboard, a desktop app for macOS, Windows, and Linux, or a Chrome extension, with every change pushed to every client in real time.
Agents connect in two ways. Claude Code attaches directly as a Claude Channel through a per-workspace MCP URL and receives task notifications natively. Codex, Antigravity, Gemini CLI, and other Agent Client Protocol (ACP) agents connect through an ACP gateway that bridges ACP to MCP so they receive the same notifications. A separate account-level Supervisor MCP server, authenticated with OAuth2, lets one agent list, create, and manage every workspace, task, and statistic in the account. Workspaces, tasks, memory, skills, and events are all exposed as MCP tools, and a CLI exposes the same workspace tools as shell commands.
Coordination features sit on top of the queue. Tasks can carry a launch date or a recurring cadence, and a background poller spawns them when due. Events are named signals that any task can fire on completion, and a Workflow arranges events and workspaces on a drag-and-drop graph so one workspace's completion creates a task in another. A workspace can be forked so a second agent takes part of its queue in its own git worktree, then merged back. Skills in the Claude Code SKILL.md format belong to the account, can be imported from any public GitHub repository, and are read by every agent that connects to a workspace where they are enabled.
Enrolled machines extend the control plane to a builder's own computers: a daemon registers a machine with a single-use code, after which the dashboard can start an agent in a workspace's folder on that machine, attach a live terminal to the running session, and show CPU, memory, uptime, and free disk. Auto-titling and speech-to-text run as in-browser models, so neither leaves the device. The backend is Go with a REST API, an SSE-based MCP server, SQLite storage, and Google OAuth2 sign-in, and the frontend is Vue 3, which the Electron desktop app renders unchanged.
Features
- MCP task tools: agents create tasks, dequeue the next assigned task, update status, reply with attachments, read and write workspace memory, publish events, and search and load skills
- Supervisor MCP: an account-level, OAuth2-protected MCP server with tools, resources, and prompts for one agent to manage every workspace and task
- ACP gateway: bridges Agent Client Protocol agents such as Codex, Antigravity, and Gemini CLI to AgentRQ's MCP notifications
- Task board: list and Kanban views of every task an agent creates, with a lane-grouped history of every tool call and message in a run
- Scheduling and events: launch dates, recurring cadences, and named completion events that spawn tasks in subscribed workspaces
- Workflows: a graph editor where dragging a workspace onto an event subscribes it and dragging an event onto a workspace emits it on completion
- Workspace forks: a fork gets its own queue and agent in a separate git worktree, shares memory, skills, and settings, and merges back when done
- Skills: SKILL.md playbooks imported from public GitHub repositories, owned by the account and enabled per workspace
- Machines: enrolled computers run agents in a workspace's folder, with live browser terminals and CPU, memory, and disk readouts
- Clients: web dashboard, Electron desktop app with native notifications and extensions, Chrome extension that bridges WebMCP sites, and a workspace CLI
- Human controls: permission requests for sensitive actions, message send delay with cancel, model selection where the agent supports it, and agent slash commands
- On-device helpers: auto-title generation and Whisper speech-to-text run as in-browser models with no server call
