MCP entry point for the MCP-CodexGPT browser bridge. Provides deterministic ChatGPT Web control built on browser-use's CDP session.
Server defines 5 tools with explicit schemas and descriptions. Tool names follow verb_noun pattern (chatgpt_status, chatgpt_ask, etc.). Descriptions are detailed and contextual, ranging 150-800+ chars, well above the 10-1024 baseline. However, 4 of 5 tools have empty input schemas ({}), violating the requirement that every parameter needs a description. Only chatgpt_ask has a proper schema with typed parameters and descriptions. Error handling is present but generic (returns ok/error/error_type). No output schemas documented. No tool annotations (readOnlyHint/destructiveHint). Composition is reasonable, tools are single-purpose and chainable (status → ask → last_response). Security: no credentials exposed as parameters; bridge manages browser session internally.
Send one prompt and wait for its confirmed final response. Use after chatgpt_status returns ``ready``. Do not send another prompt or call another browser tool while waiting. Progress notifications and response tails are context, not final advice; progress values do not decrease when ChatGPT temporarily hides an already-started response turn. Accept only ``ok=true`` with ``status="completed"``. By default, post-submit rendered-text mismatch is advisory because ChatGPT's UI representation can change; exact pre-click composer verification plus conversation, target, and user-turn-count ownership remain mandatory. Such a completion returns ``correlation_status="rendering_fallback"`` and ``correlation_warning``. Set ``strict_user_turn_text=true`` only when a rendered-text mismatch must fail. After a timeout, recover with chatgpt_last_response without resending. After any ambiguous non-timeout error, discard partial output and do not retry; a new attempt requires explicit user direction.
Close ChatGPT after tool work or an already returned failure. Never close while chatgpt_ask is waiting. Finish any allowed timeout recovery first; cancellation after submission is not a rollback.
Read the latest response after chatgpt_ask explicitly times out. Do not resend the prompt. ``in_progress`` is provisional; accept only ``completed``. Do not use this after an ambiguous non-timeout error because the latest response may belong to another turn.
Open a fresh conversation when no chatgpt_ask call is running. For a normal independent request, prefer ``chatgpt_ask(new_chat=True)``. Never use a new chat to hide or recover an ambiguous post-submit error.
Four tools (chatgpt_status, chatgpt_close, chatgpt_new_chat, chatgpt_last_response) have empty input schemas ({}). This violates pattern:tool requirement that every parameter needs a description.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract fields without knowing response structure. chatgpt_ask returns dict with ok/status/correlation_status/correlation_warning, but this is only visible in docstring, not in a formal schema.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. chatgpt_status and chatgpt_last_response are read-only; chatgpt_close, chatgpt_new_chat, and chatgpt_ask are destructive/write operations. Annotations help agents reason about side effects and retry safety.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Open ChatGPT and wait for one final readiness result. Use once before chatgpt_ask. Wait for this call; the bridge owns the bounded readiness loop and never returns transient ``loading``. Continue only at ``ready``; ``login_required`` or ``challenge`` needs user action. Do not call any browser tool while chatgpt_ask is waiting.
Error handling returns generic {ok: False, error: str, error_type: str}. Does not categorize errors as retryable, user-fixable, or fatal. Does not provide recovery guidance (e.g., 'login_required' should suggest calling chatgpt_status again after user signs in).
chatgpt_ask timeout_seconds parameter accepts 10-900 but no validation shown in code. If LLM passes 5 or 1000, error handling does not guide correction. Parameter description mentions constraint but code does not enforce it.