Persistent browser daemon + MCP server + agentic loop that routes prompts to chatgpt.com — no API key required
Three tools with basic schemas and descriptions, but significant gaps in parameter documentation and error handling. chatgpt_ask has a well-structured inputSchema with 7 optional parameters, each with descriptions (good baseline). However, no output schema is documented, LLMs cannot predict what fields to expect from ChatGPT's response. chatgpt_status and chatgpt_stop have empty inputSchemas and minimal descriptions (under 20 chars for status). No error recovery guidance, no validation rules, no idempotency hints. Tool names are verb-noun (good), but descriptions lack WHEN/WHY context. No mention of prerequisites, failure modes, or what the response structure contains.
Ask ChatGPT a question via chatgpt.com using a persistent browser. Useful as a second opinion or when the main model's knowledge may be stale. The daemon auto-starts on first use and stays alive between calls.
Check whether the ChatGPT browser daemon is currently running.
Shut down the ChatGPT browser daemon and close the browser.
No output schema documented for any tool. LLMs cannot predict response structure (fields, types, nesting). chatgpt_ask returns text via ChatGPT, what format? Markdown? JSON? Plain text? Undocumented.
chatgpt_status and chatgpt_stop have empty inputSchemas (type: object, properties: {}) and descriptions under 20 characters.
No error handling guidance. If ChatGPT is unreachable, authentication fails, or the daemon crashes, the tool returns raw text. No recovery hints (e.g., 'Try chatgpt_status() first' or 'Restart with chatgpt_stop() then retry').
Parameter 'file' accepts absolute paths with no validation or sanitization. No mention of path traversal checks, file size limits, or supported MIME types. LLMs could pass malicious paths.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
No idempotency or deduplication semantics documented. The mcp-server.js implements a 60-second cache to prevent duplicate ChatGPT calls, but this is invisible to the LLM. If the agent retries, it may get stale cached results without knowing.