agentrq/agentrq

AgentRQ – A Collaboration Platform for Humans + AI Agents

What it is – AgentRQ is an open‑source, self‑hostable web/desktop application that lets a human operator and one or more AI agents (e.g., Claude, Antigravity, Codex) share a common task workspace. The platform exposes a Model Context Protocol (MCP) server that agents can call as if it were a set of remote tools: create tasks, update status, reply to messages, fire events, read/write “memory” notes, etc. Because the agents see the same state as the human UI, they can work autonomously while the user stays in the loop.

---

Core Features (as described in the README)

Feature	What you get
Visual Task Board	Real‑time Kanban/list view of every task an agent creates, with status, requirements and history.
Task Scheduling	Assign a launch date or recurring cadence (15 min‑hour‑day‑week‑custom). A background poller spawns the task exactly when due – no server‑side “always‑on” agents required.
Events & Workflows	Named signals (qa_passed, deploy_finished, …) that can trigger new tasks in other workspaces. Drag‑and‑drop graph editor builds event‑to‑workspace pipelines without YAML or DSL.
Tool‑Call History	Timeline view of every tool call and chat message in a task run, searchable and drill‑downable.
Auto‑Title Generation	One‑click title creation using a tiny on‑device language model (cached in the browser, no API calls).
Speech‑to‑Text	In‑browser Whisper transcription for task descriptions or replies – processed locally, never uploaded.
Message Send Delay	Optional countdown (3‑60 s) before a message reaches the agent; can be cancelled or sent immediately.
Keyboard‑Driven UI	Global shortcuts (⌘K/Ctrl+K task finder, N new task, M/T chat/trajectory view, ? help) for fast, mouse‑free work.

---

Architecture (high‑level)

• Backend – Go (Fiber) REST API + an SSE‑based MCP server (mcp-go). Data stored in SQLite via GORM, authentication via Google OAuth2 + JWT, internal pub/sub for real‑time updates.
• Frontend – Vue 3 + Vite, state managed with Pinia, styled with Tailwind CSS and a glass‑morphism UI.
• Desktop client – Electron wrapper that runs the same Vue UI, adds native notifications, tray menu, global shortcuts, deep‑link handling (agentrq://…) and auto‑updating.
• Extensions – Desktop‑only Node modules discovered on GitHub (agentrq‑extension topic). Extensions can add pages, actions, shortcuts, or scheduled jobs; they run with the user’s privileges and are mediated by the core app (no server‑side code execution).

---

How an AI Agent talks to AgentRQ

1. MCP URL & token – Each workspace publishes a unique MCP endpoint (https://<workspace>.mcp.agentrq.com/?token=…).
2. .mcp.json – Placed in the project directory; tells the agent which MCP server to use.
3. Claude integration – A .claude/settings.local.json file can pre‑approve the set of MCP tools (createTask, reply, loadMemory, etc.) so Claude Code can act without prompting the user each time.
4. ACP Gateway – For agents that speak the Agent Client Protocol (Antigravity, Codex), @agentrq/acp-gateway bridges ACP ↔ MCP, handling authentication, task routing and real‑time notifications.

---

Getting Started (quick‑start steps from the README)

1. Prerequisites – Go 1.21+, Node 18+, a Google OAuth2 client ID/secret.
2. Clone & install

   git clone https://github.com/agentrq/agentrq.git
   cd agentrq
   make install          # pulls Go modules, npm packages, etc.
   

3. Run locally

   make dev               # starts backend (Fiber) and frontend (Vite) together
   # in another terminal for the desktop UI
   make desktop-dev
   

   UI will be reachable at http://localhost:5173.
4. Desktop builds – Pre‑built installers are available on the Releases page, or install via the one‑liner:

   curl -fsSL https://agentrq.com/install.sh | sh -s -- --quit
   

5. Self‑hosting – Docker image and SETUP.md provide a production‑ready compose file.

---

Typical Use Cases

• AI‑assisted project management – Let Claude generate, schedule, and update tickets while the human reviews and provides occasional clarification.
• Continuous‑integration pipelines – Emit events (deploy_finished) that automatically spawn post‑deployment verification tasks for an AI reviewer.
• Knowledge‑base building – Agents write to memory.md (workspace notes) that persist across runs, enabling cumulative learning.
• Hybrid coding assistance – Codex or Antigravity can be hooked into the same workspace, receiving task assignments and replying with code snippets, all visible in the shared dashboard.

---

Where to Learn More

• Full feature list: https://agentrq.com/features
• Desktop guide: docs/DESKTOP.md
• Extension development: docs/EXTENSIONS.md
• MCP protocol spec: linked from the repo (WebMCP integration)
• ACP gateway docs: docs/WEBMCP.md, docs/SLASH_COMMANDS.md

---

Bottom line – AgentRQ is a concrete, production‑grade platform that bridges modern LLM agents with a human‑friendly task board, exposing a well‑defined tool protocol (MCP) for autonomous yet supervised AI work. It is fully open‑source, self‑hostable, and includes both web and native desktop clients.