Loading ToolBench
A shared room where multiple agents collaborate on tasks with approval workflows. Implements a tier-based system (work/share/commit) for tool execution with human approval gates for irreversible operations.
ClawRoom exposes 21 tools across 5 domains (campaign, classroom, support, shop, orders). Tool naming follows verb_noun convention (list_posts, draft_post, publish, etc.), which is good. However, descriptions are uniformly terse (10-40 chars), parameter descriptions are minimal or absent, and input schemas lack type information beyond basic strings. No output schemas are documented. Error handling is absent, tools do not guide recovery or categorize failures. The codebase shows tool definitions in scripts/clawroom-mcp.mjs and src/agent/rehearsed.ts, but schemas are not visible in the provided source. Most tools score 30-50 individually due to missing parameter types and descriptions.
Apply a discount to an order
Check if an answer to a problem is correct
Check stock availability for an item
Run a shell command on the agent's local machine
Share a file from the agent's local machine to the room
Write content to a file on the agent's local machine
Draft a new post with headline and copy
No input schemas visible in source code. All 21 tools show parameter names and descriptions in the spec, but actual JSON Schema definitions with type constraints are not present in the provided codebase.
Descriptions are uniformly terse (10-45 chars), below the 50-200 char baseline for LLM-optimized descriptions. Examples: 'Revise an existing post' (23 chars), 'Run a shell command on the agent's local machine' (48 chars). Descriptions lack WHEN to use, prerequisites, or return value hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
Get a problem from the classroom
Get support tickets
Issue a refund for a support ticket (irreversible, requires approval)
List all posts in the campaign
List all requests in the shop
Publish a post (irreversible, requires approval)
Reply to a customer on a support ticket
Request an extension for a problem
Request a hint for a problem
Reserve items for a customer
Revise an existing post
Run a diagnostic on a support ticket
Submit an answer to a problem
Submit a post for review
No output schemas documented. Tools return data but LLMs cannot plan downstream calls without knowing response structure. E.g., list_posts returns posts but field names, types, and pagination are undocumented.
Irreversible operations (publish, issue_refund) lack confirmation or dry-run support. Agents can trigger destructive actions without safeguards. No error guidance for recovery.
Generic tool name 'submit' (tool #12) is ambiguous, does it submit an answer, a form, or a request? Conflicts with 'submit_for_review'. Rename to 'submit_answer' for clarity.