Skip to main content

​
Connection and setup

  • 405 Method Not Allowed or SSE connection failure: use Streamable HTTP. In Claude Code, set type to http or streamable-http, not sse. The data endpoint accepts POST /mcp.
  • Claude’s Add custom connector button is missing: use Customize > Connectors for personal plans, or Organization settings > Connectors for Team and Enterprise owners.
  • ChatGPT cannot add a connector: use ChatGPT on the web and check your plan and workspace permissions before enabling Developer Mode. See the plan-specific requirements below.
  • Tools do not appear after editing configuration: restart the AI client. Some tools need the admin toolset; others require additional scopes.
See setup for complete client instructions.

​
ChatGPT plan and workspace requirements

OpenAI’s MCP apps help article distinguishes these cases:
  • Personal Pro: enable Developer Mode in your own settings. The help article describes read/fetch MCP access, not the full write-action beta. Personal accounts do not need workspace-admin approval.
  • Business: an admin or owner must enable Developer Mode for themselves and create and publish the custom app under Workspace settings > Apps. Ordinary members use the published app.
  • Enterprise/Edu: an admin or owner grants Developer Mode access through Permissions & Roles > Connected Data. Authorised members can then enable it in their own settings; only admins/owners publish apps and control workspace access.
Full MCP support, including write actions, is described in that help article for Business and Enterprise/Edu. OpenAI’s developer guide lists broader plan eligibility and a newer Settings > Security and login > Developer mode path, while the help article uses Settings > Apps > Advanced Settings. Availability and menus are changing: check the controls shown in your account rather than assuming every plan exposes the same tools.

​
Authentication and permissions

  • Missing Authorization header or expired OAuth token: reconnect Redbark from the client to complete OAuth again.
  • Invalid or revoked API key: create a replacement in Settings > API & MCP.
  • plan_upgrade_required: this is a plan gate, not a credentials problem. Check plan requirements and billing.
  • Consent screen shows Upgrade to connect: your plan does not include MCP. Choose View plans to upgrade, then connect again. Cancel returns access_denied to your AI client with the plan reason. An access_denied without that reason can also mean you declined consent.
  • insufficient_scope or a missing named scope: disconnect and reconnect the client, granting that permission. Brokerage reads require brokerage:read as well as an eligible plan.

​
Invalid arguments

Rejected arguments return an invalid_arguments tool result naming the problems. An unknown argument also lists accepted names. Refresh the client’s tool schema and correct the named argument before retrying. Use the checked examples as a guide.
  • Copy public ids from discovery responses. Do not substitute names, account numbers, or database UUIDs.
  • Use account for balances, details, transactions, holdings, and trades. get_account uses id; list_accounts uses connection to filter.
  • Use the opaque page token for pagination, not offset or the whole next-page URL.
  • Check account eligibility. Banking balance tools cannot read brokerage or Documents accounts.
  • Keep from on or before to, and use supported date formats and real calendar dates. For example, 2026-02-30 is rejected rather than moved into March.
Repeating the unchanged call will not fix validation, an unsupported account type, or a missing workflow prerequisite.

​
Rate limits

Transaction, holding and trade reads share a per-minute budget and a cap on how many can run at once. search_transactions and spending_summary make one read per account, plus one more for every 1,000 transactions an account has in the window, so several of them at once can use up the budget.
  • The server paces these reads. It queues them under the cap, waits out a short limit and retries, so most overlapping calls finish without an error.
  • A paced call can take up to 30 seconds longer than usual.
  • If a call still hits the limit, the error is rate_limit_exceeded or too_many_concurrent_requests. It usually says how many seconds to wait. Wait that long, then retry.
  • To use fewer reads, run one search at a time, pass accounts, or narrow from and to.
You do not need to reconnect or create a new API key for a rate limit.

​
Provider waits

provider_wait means a banking provider read has been deferred, not that your API key has used its allowance. The server retries GET reads when this error has a Retry-After of at most 5 seconds, with up to six retries inside the same 30-second waiting budget. Writes, longer waits and waits without Retry-After are not retried automatically. If the tool still returns provider_wait, follow its retry delay. Reconnecting or replacing your API key will not clear a provider cooldown.

​
Getting help

When an API error includes a request id, quote it to support along with the tool name and error code. Do not include banking payloads or access tokens. For error formats and endpoint details, see the API v2 error reference.