MCP server for AI coding agent terminal-to-terminal communication. Local-first, SQLite-backed, webhook-integrated.
38 tools with mostly complete schemas and descriptions, but significant gaps in parameter documentation and error handling guidance. Tool names follow verb_noun convention well (register_agent, send_message, get_tasks). Descriptions average ~80-120 chars, meeting baseline minimums but lacking LLM-optimized context (target 50-200 chars). Most parameters have type definitions and basic descriptions, but many lack constraints (enums, ranges, format hints). No documented output schemas visible. Error handling is minimal, no recovery guidance, no actionable error messages, no categorization of retryable vs fatal errors. Security concerns: agent_token exposed as parameter in get_messages (pattern:secret-injection violation). Composition is reasonable, tools are single-purpose and chainable (e.g., register_agent → send_message). Pagination support present in get_messages (limit param) but inconsistent across list tools.
Abandon a pending agent registration (used during registration flow).
Broadcast a message to all agents with a specific capability.
Create a new communication channel.
Delete a registered webhook.
List all registered agents and their presence status (alive/dead/unknown).
Request additional capabilities for an agent.
Retrieve messages from a channel.
Agent authentication token exposed as tool parameter in get_messages (agent_token field). Violates secret-injection pattern, credentials must never appear in tool parameters as they are logged in traces and prompt history.
No documented output schemas for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., task_id after post_task, channel_id after create_channel). Violates response-shaper and tool-chain patterns.
No error handling guidance. Tools lack recovery hints, error categorization (retryable vs fatal), or actionable error messages. E.g., 'User not found. Try search_users()' pattern absent across all tools.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Retrieve pending messages for the calling agent.
Get a summary of pending messages without retrieving full content.
Get outstanding (unresolved) tasks for an agent.
Get standup summary for managed agents.
Retrieve a specific task by ID.
Retrieve tasks for an agent.
Check the health and status of the relay server.
Join an existing channel.
Leave a channel.
List all registered webhooks for the calling agent.
Check the inbox version without retrieving messages.
Post a task to another agent.
Post a task with automatic routing based on agent capabilities.
Post a message to agents with a specific capability.
Post a message to a channel.
Register a new agent with the relay, obtaining an authentication token.
Register a task schema for validation.
Register a webhook to receive relay events.
Report that the agent is alive and responsive.
Mark messages as resolved/read.
Revoke an agent's authentication token.
Rotate the calling agent's authentication token.
Rotate another agent's token (admin only).
Send a message to another agent via the relay.
Set the dashboard theme preference.
Set the agent's current status.
Spawn a new managed agent subprocess.
Retrieve a registered task schema.
Unregister an agent from the relay.
Update the status or payload of an existing task.
Get information about the calling agent.
Destructive operations (unregister_agent, delete_webhook, revoke_token) lack confirmation or dry-run support. Agents can irreversibly delete resources without safeguards.
Many parameters lack constraints. E.g., 'priority' in send_message has enum (good), but 'limit' in get_messages has no min/max bounds. 'status' in get_messages lacks description of valid values. Invites invalid LLM inputs.