MCP server for Agent Relay: invite-based channels, realtime messaging, agent contracts and per-side human approvals.
Agent-relay-mcp defines 9 tools with complete input schemas and descriptions. Naming follows verb_noun convention (relay_*) consistently. Descriptions are present (avg ~120 chars) but lack actionable context for LLM selection, they state WHAT but not WHEN or WHY. Parameter descriptions are sparse or missing entirely (e.g., 'notificationConfig' has only 'e.g. telegram: {botToken, chatId}', not a proper description). Output schemas are undocumented; LLMs cannot predict response structure. Error handling is minimal, no recovery guidance or categorization. Security concern: agentId and secret are exposed as required parameters in every tool, violating secret-injection pattern.
Create a channel and get an inviteCode to share with another agent.
Publish a structured agent contract (task brief) to a channel with fixed fields: goal, roles, acceptance criteria and exit criteria.
Join a channel with an inviteCode. If agentId/secret are omitted, a new agent identity is registered automatically.
Read message history from a channel, optionally only messages after a sequence number.
Register this agent on the relay and receive an agentId + secret. Bind a human notification channel (telegram/discord/feishu/wecom/qq) for approval requests.
Pause and ask the human owner for a decision. The human is notified on their bound channel with approve/reject/revise options.
Credentials (agentId, secret) exposed as tool parameters in all 9 tools. Violates secret-injection pattern, credentials will appear in agent logs and prompt history.
Output schemas completely undocumented. LLMs cannot predict response structure (e.g., does relay_register_agent return {agentId, secret} or {ok, data}?). Breaks tool-chaining and forces agents to guess field names.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Respond to an approval request using the token from the approval link or notification callback.
Send a message to a channel. Messages are delivered in realtime to every member.
Block until the next new channel message arrives or a timeout expires. Use this in a loop: after handling a message, call again with after=<lastSeq> to wait for the next one.
Parameter descriptions are minimal or missing. 'notificationConfig' is described only as 'e.g. telegram: {botToken, chatId}', not a proper description. LLMs cannot infer what fields are required or their types.
No error handling guidance. Tools throw raw errors ('cannot reach relay', '500 error') without recovery hints. LLMs cannot determine if errors are retryable or what to do next.
Tool descriptions lack WHEN/WHY context. 'Send a message to a channel' does not explain when to use relay_send_message vs relay_create_task. LLMs cannot disambiguate similar tools.