AgentRQ integrates with Claude Code via the native MCP protocol. One entry in .mcp.json gives Claude access to 18 purpose-built tools for
human
collaboration — no wrappers, no hacks, no polling loops.
{
"mcpServers": {
"agentrq": {
"type": "http",
"url": "https://WORKSPACE_ID.mcp.agentrq.com/mcp?token=TOKEN"
}
}
}Each tool serves a specific purpose in the Claude ↔ Human collaboration loop.
createTask
Claude's primary way to ask for your input. Creates a task with a title, full context in the body, and optional file attachments. You see it instantly in the dashboard.
updateTaskStatus
Transitions a task through its lifecycle: notstarted → ongoing → completed / rejected. Call immediately when starting work, and when done. Keeps the dashboard accurate.
reply
Sends a message in a task thread — from Claude to you. Use it for progress updates, follow-up questions, or sharing output (with file attachments like diffs or logs).
getWorkspace
Fetches workspace metadata and the mission context you've set. Call at the start of every session to confirm connectivity and load any project-specific instructions.
getTask
Pulls the next priority task from the queue — or a specific task by ID — and can include its full conversation history. Perfect for autonomous agents that work in loops and for resuming long tasks without polling the dashboard.
getAttachment
Fetches a file you attached in the dashboard: its public link by default, or its content as base64. Enables true bidirectional file exchange — you can hand Claude a design, config, or dataset mid-session.
publishEvent
Fires a named signal when a stage finishes. Every workspace subscribed to that event spawns its trigger task automatically — this is how one agent's output becomes the next one's input.
elicit
Asks you a question and blocks until you answer — a form to fill in, or a link to open and confirm. The agent waits for a real decision instead of guessing and carrying on.
loadMemory
Reads what this workspace remembers from earlier tasks. The memory belongs to the workspace, not the agent, so a different model on a different machine starts with the same notes.
saveMemory
Writes something worth remembering, so the next task starts with it instead of asking you again.
MEMORY.md is the index; detail lives in named memories
linked from it.
deleteMemory
Deletes one of the workspace's memories by name. Deleting a name nobody wrote under — or one already deleted — is not an error, it just says there was nothing to remove.
searchSkills
Finds the skills this workspace can use — its own and those shared into it — by name or description. Only descriptions come back, never bodies, so having many skills stays cheap.
loadSkill
Reads one file of a skill by its skill:// URI. A
SKILL.md comes back with the URIs of the skill's other files, loaded only when it points to them.
saveSkill
Writes one file of one of this workspace's own skills, replacing it completely. Writing a
SKILL.md creates the skill if it does not exist yet.
deleteSkill
Deletes one of this workspace's skills, or one of its files. Skills shared in from another workspace are read-only here, and deleting something already gone is not an error.
listSiteTools
Lists the websites you shared with the workspace from the AgentRQ Chrome extension, and the name and description of each WebMCP tool they offer, with a flag for whether your Chrome is online.
getSiteToolDefinition
Returns one shared site's tool in full, with its input schema and annotations. Read-only; the agent calls it before callSiteTool, so the listing stays small.
callSiteTool
Runs a shared site's tool in your own Chrome, signed in as you. Anything the site does not mark read-only waits for your Allow once, Always allow or Deny in the task.
The most powerful way to use AgentRQ is to add instructions to your CLAUDE.md — telling Claude exactly when to ask you, what
context
to include, and what it can proceed with autonomously.
## AgentRQ — Human-in-the-Loop
At the start of every session:
→ Call `getWorkspace` to load mission context.
→ Report what you're connected to.
Before any of these, STOP and call `createTask`:
- Database schema changes or migrations
- Anything that deletes or drops data
- Production deployments or config changes
- External API calls with write side effects
- Spending money (OpenAI, Stripe, etc.)
When creating a task, always include:
- What you've done so far
- Exactly what you need approval for
- The risk if they say yes
- The fallback if they say no
After human replies via channel:
→ Read the message carefully
→ Call `updateTaskStatus("completed")`
→ Continue with their instructions
# Be specific in tasks. Vague asks = slow unblock.No proprietary protocols. No vendor lock-in. AgentRQ uses standard MCP SSE transport — the same protocol powering the Claude ecosystem.