Local relay MCP and dashboard for ChatGPT Workspace Agent trigger callbacks.
Four tools with complete schemas and detailed descriptions. Tool naming follows verb_noun pattern (record_*, server_info). Descriptions are comprehensive (200-400 chars), exceeding baseline averages. All parameters have types and descriptions. However, output schemas are not explicitly documented in the source, and error handling guidance is minimal. Tool annotations present (readOnlyHint, destructiveHint, idempotentHint). No security issues detected (no credentials in params). Composition is sound, each tool has a single responsibility in the relay workflow.
Use this at the START of a turn to share your step plan with the local operator. The operator cannot see your ChatGPT-side plan, so this is their only way to know what you are about to do. Each step needs a stable id (reuse the same ids in record_progress step_updates) and a short title. Use user-visible work steps; do not include relay binding, server_info, or routine tool setup unless that plumbing is the task. Call it again to REVISE your plan when the direction changes — a plan revision is NOT a finished turn, so do not call record_result for it; instead skip the old steps via record_progress or replace them here. Returns the current plan and run status. Example: steps=[{"id":"s1","title":"Confirm changed files"}, {"id":"s2","title":"Fix Header sizing"}, {"id":"s3","title":"Run build"}]
Use this after completing a chunk of work to batch-sync plan step statuses back to the operator. Pass step_updates to flip one or more step ids to a new status (done/in_progress/skipped) — use skipped to abandon a step when the direction changes, instead of ending the turn. Optionally include a one-line message summarizing what you just did. Do not call this once per tiny tool call — batch several completed steps into one call. Unknown step ids are ignored and listed in the returned ignored_step_ids. Returns the updated plan and run status. Example: step_updates=[{"id":"s1","status":"done","note":"boundaries confirmed"}, {"id":"s2","status":"in_progress"}], message="Boundaries done, fixing Header"
Use this EXACTLY ONCE when this turn is truly over. Records the final status, a title, and the full Markdown answer. This is what the operator reads as the result — do not only answer in the ChatGPT conversation. status: done when delivered, failed on an execution error, blocked ONLY for an external hard blocker (missing access/resource/dependency) — never use blocked to mean 'the plan changed' (revise the plan with record_plan / step_updates instead). Attach text artifacts if useful. Returns the final run status. Example: status="done", title="UI fixes applied", markdown="## Summary\n...", artifacts=[{"name":"diff.md","mime_type":"text/markdown","content":"..."}]
Output schemas not documented in source code. While input schemas are complete with types and descriptions, return types for all four tools are not explicitly defined. LLMs cannot plan downstream operations without knowing what fields to expect.
Error handling lacks recovery guidance. Tools do not document what errors are retryable, user-fixable, or fatal. No examples of error responses or actionable error messages provided.
No pagination or result limits documented. record_plan accepts 1-20 steps (good), but server_info and other tools do not specify limits on returned data or pagination strategy.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | 2026-07-28+ | v2 |
Use this to orient yourself: returns the relay app name, version, state paths, auth mode, and the list of tool names available on this server. Read-only.
Artifact handling in record_result lacks validation constraints. The artifacts array accepts arbitrary mime_type and content without documented limits on size, count, or supported formats.
request_id and conversation_key parameters lack format documentation. No guidance on expected format, length, or character restrictions provided.