Codex desktop: disclose Robinhood MCP execution restrictions before workflow setup and clarify human-approved orders

Summary

Submitting at the affected user's explicit request. In Codex desktop on macOS, an assistant helped plan and build a Robinhood Agentic investment-committee workflow, including a shadow period and a proposed human-approval-gated live pilot. Only when real execution was requested did it clearly state that it cannot submit securities orders even after explicit per-trade approval. Manual brokerage submission was then offered as a substitute.

The user considers this late-disclosed limitation an unacceptable product experience. The request is for an authoritative supported-behavior answer and earlier capability disclosure, not a way to bypass safeguards.

Environment and observed boundary

• Standalone Codex desktop app on macOS, persistent task, Robinhood MCP integration.
• Report date: September 18, 2026.
• Exact app build and selected model ID are not independently captured in this report; no controlled cross-model comparison was performed.
• Refusal occurred in conversation before an order-submission tool call. This is not a diagnosed broker rejection, OAuth failure, or rejected approval card.
• Paraphrase of refusal, not a verbatim log: the assistant cannot submit securities trades on the user's behalf, even with explicit approval of the exact order.
• No live trade was attempted to reproduce this report.

Expected behavior

1. Clearly disclose execution limitations before helping a user invest substantial effort in a workflow dependent on execution.
2. Distinguish tool availability, brokerage authorization, application permissions, and assistant-level restrictions.
3. State whether exact, per-transaction human approval can support order submission in Codex, or whether the human must perform the final broker action.
4. If approved execution is supported, document the supported configuration. If prohibited, explicitly document that limit for Codex and third-party trading integrations, and consider supporting a bounded human-approved flow.
5. Do not imply a manual-execution substitute fulfills an agent-execution request.

Documentation mismatch requiring clarification

Robinhood's official documentation lists Codex/Codex CLI compatibility and describes agent order placement:

• https://robinhood.com/us/en/support/articles/agentic-trading-overview/
• https://robinhood.com/us/en/support/articles/trading-with-your-agent/

Broker-side capability does not establish OpenAI permission. However, the distinction and practical limitation were not clear to the user until late in setup.

Prior support attempt

The official Help Center chat was contacted. Its automated assistant cited generic app-action restrictions, said it could not transfer to a human or issue a case number, and directed the user to this tracker. Please route to the appropriate Codex product/policy owner or identify a private escalation route.

Related reports

• #44163: authorized securities commitment and approval-boundary clarification.
• #44554: another user's report of session-dependent Robinhood refusals.

Those reports are context, not verified reproductions by this reporter or evidence that changing models/sessions is authorized. This issue emphasizes upfront capability disclosure, misleading workflow expectations, and the support-escalation gap.

No email address, account identifiers, holdings, credentials, private project names, or private conversation exports are included.

Start with the issue's cited Robinhood documentation and related reports #44163 and #44554; no repository file or test is identified. Establish Codex's authoritative supported behavior for Robinhood execution and human approval, then document the limitation and when it must be disclosed before workflow setup.