Loading ToolBench
An MCP server that enables inter-agent communication for Claude Code instances working on the same project. Agents can discover each other, send messages, and coordinate work to avoid conflicts.
Server defines 6 tools with clear, action-oriented names (who, send, reply, peek, ack, ack_all) and detailed descriptions (avg 180 chars). All tools have Zod schemas with typed parameters and descriptions. However, output schemas are not formally documented, responses are inferred from code. Error handling is minimal (e.g., 'Message not found' is bare). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Parameter descriptions are good but lack format constraints (e.g., message_id format, message length limits). Composition is sound, each tool has one responsibility and chains via agent codes.
Mark a message as read and remove it from your inbox. Use after you've processed a message that doesn't need a reply.
Clear your entire inbox. Use when you've read all messages and none need a reply.
Check your inbox for messages from other agents. Call this at the START of any task and PERIODICALLY during long work. If you have messages, read them and reply or acknowledge before continuing.
Reply to a specific message. Sends your response to the original sender and removes the original from your inbox. Always reply to messages that ask you a question or request information.
Send a message to another agent or broadcast. Use 'all' as target to broadcast to all agents on the same project. Use a specific code (from 'who') to message one agent directly. Proactively send messages when: - You're about to modify a file that others might be editing - You need to coordinate a PR, merge, or deployment - You want to ask what others are working on to avoid conflicts - You finished a task that unblocks someone else
Output schemas not formally documented. Responses are inferred from code (text content arrays). LLMs cannot plan downstream operations without knowing return structure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools have clear risk profiles (send/reply/ack are WRITE; who/peek are READ_ONLY) but LLMs cannot infer safety without explicit hints.
Error handling lacks recovery guidance. 'Message not found' and 'not found in inbox' are bare, no suggestion for next steps (e.g., 'Call peek() to refresh inbox').
Parameter constraints missing. message_id and message parameters lack format/length constraints. LLMs may pass oversized or malformed IDs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
See which other Claude Code agents are currently active. Shows their code, project, and uptime. Use this BEFORE starting work to check if other agents are working on the same project. Use this to find an agent's code before sending them a message. By default shows only agents on the same project. Set scope='all' to see everyone.
No pagination or result limits documented. peek() returns all messages without limit, large inboxes could blow context window.