Loading ToolBench
AgentRQ is an AI-powered task queue for autonomous agents with workspace-based orchestration, multi-agent coordination, human-in-the-loop approval, and MCP protocol integration.
AgentRQ exposes 17 tools with generally clear naming and descriptions. All tools have non-empty descriptions (baseline: 194 chars average). Most parameter inputs are typed with descriptions. However, output schemas are completely undocumented, the JSON source provides tool names, descriptions, and input schemas but zero output schema documentation. This is a critical gap for production use. Tool names follow verb_noun convention well (get*, list*, create*, update*, archive*, unarchive*, delete*). Parameters are mostly well-described with context. One tool (agentrq_autopull) has an enum-constrained action parameter (status|pause|resume|pull_now), showing schema discipline. Risk annotations are present (READ_ONLY, WRITE, DESTRUCTIVE), which is good. However, some tools lack sufficient granularity in parameter constraints, and error handling guidance is not visible in the provided source.
Inspect or steer automatic delivery of AgentRQ work. Each task gets its own dedicated session, opened the first time it is pushed and closed once it reaches a terminal status. The workspace pushes tasks and messages on its own; this tool does not fetch them on a timer. "status" reports the workspace connection and whether delivery is on; "pause" and "resume" stop and restart opening/routing task sessions; "pull_now" dequeues the next task assigned to you right away and returns it in this call's result, without opening a dedicated session for it. Task content, replies, and status changes go through the mcp__agentrq__* tools, not this one.
Archive a workspace, hiding it from the default list without deleting anything.
Create a workspace. The description is the mission agents are given.
Permanently delete a workspace and everything in it. This cannot be undone — prefer archiveWorkspace unless the user has asked for deletion in so many words.
Activity statistics across every workspace the user owns, over a time range, with a per-workspace breakdown of what drove the totals.
Output schemas completely undocumented. None of the 17 tools document what fields they return, data types, or structure. This forces LLMs to guess response structures and prevents downstream tool chaining.
No pagination schema visible. listWorkspaces and listWorkspaceMemories likely return arrays but no limit/offset parameters or next_cursor field documented. Large result sets risk exceeding context windows.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | A | 82 | 2026-07-28+ | v2 |
Where the user is right now in the AgentRQ interface: the route path and its parameters. Call this first to resolve phrases like "this task" or "the workspace I am looking at" into the IDs the other tools need.
Who is signed in. Everything else these tools do is done as this user.
One workspace, including its mission description and settings.
One of a workspace's memories, in full. Start with MEMORY.md, which indexes the rest.
Activity statistics for a workspace over a time range.
The workspace token an agent uses to connect to this workspace over MCP. This is a credential: show it to the user, never paste it into anything else.
What the agents working in a workspace have written down for each other: name, size and when each was last changed. The content is not included — read one by name for that. MEMORY.md is the index the others hang off.
Every workspace the signed-in user can see.
Open a page in the AgentRQ interface, moving the user there. Paths are the ones in the address bar, for example "/", "/tasks/ongoing", "/workspaces/<workspaceId>/board", "/workspaces/<workspaceId>/tasks/<taskId>", "/events", "/workflows".
Bind a workspace's notifications to a Slack channel.
Restore an archived workspace to the default list.
Change a workspace. Only the fields given are altered.
Parameter constraints insufficiently formalized. getWorkspaceStats accepts 'range' like '7d' or '30d' but no enum constraint visible; LLMs may pass invalid ranges. Unix timestamp fields (from/to) lack min/max bounds documentation.
Error handling and recovery guidance absent from tool descriptions. No indication of what errors each tool can raise, when to retry, or which errors are user-fixable vs fatal.
getWorkspaceToken description warns to 'show it to the user, never paste it into anything else' but this is a frontend tool. Unclear if credentials leak into agent logs or prompt history. Server-side secret injection pattern not evident.
deleteWorkspace is destructive but no confirmation/dry-run pattern evident. Tool description correctly warns agents to prefer archiveWorkspace, but no confirmation-request or idempotent-operation guidance for agents prone to mistake.