Loading ToolBench
MCP-based inter-agent communication tool for Claude Code
Four tools with complete Zod schemas and descriptions. Naming follows verb_noun pattern (discover_, list_, ask_). Descriptions are substantive (100-250 chars) and explain WHEN to use each tool. Parameters are well-typed with Zod constraints (z.string(), z.array(), z.boolean()). However, output schemas are not documented, responses are JSON-stringified without field-level documentation. Error handling is basic (try/catch with generic error messages). No tool annotations (readOnlyHint/destructiveHint). ask_agent_execute is destructive but lacks confirmation pattern or dry-run option.
Send a request to another agent with execution privileges. Specify which tools the agent should have access to (built-in tools like Bash/Write/Edit and MCP tools like mcp__supersurf__browser_navigate). Use discover_agents then list_agents to find available capabilities. You can pass individual tool names via "tools" and/or entire MCP server names via "servers" to grant access to all tools from those servers.
Send a read-only question to another agent. The target agent can read its project to answer but cannot modify files or run commands. Optionally attach files from your project for the target agent to review.
Discover available dialup agents. Returns a lightweight directory of agent names and descriptions. Use the optional filter parameter to search for agents with specific capabilities (e.g. "supersurf", "Bash").
Get full capabilities for specific dialup agents. Returns tools and MCP servers available on each requested agent. Use discover_agents first to find agent names.
Output schemas not documented. Tools return JSON-stringified responses without field-level documentation. LLMs cannot plan downstream calls or extract specific fields reliably.
ask_agent_execute is destructive (WRITE risk) but lacks confirmation pattern or dry-run option. Agents can execute arbitrary commands on remote agents without preview or rollback.
Error messages are generic ('Error: <message>'). No recovery guidance or error classification (retryable vs user-fixable vs fatal). LLMs cannot determine next steps on failure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). ask_agent_readonly and ask_agent_execute should declare their safety properties explicitly.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
Parameter 'filter' in discover_agents and 'tools'/'servers' in ask_agent_execute lack format constraints or examples of valid values. LLMs may pass invalid capability names.