> ## Documentation Index > Fetch the complete documentation index at: https://docs.backquant.com/llms.txt > Use this file to discover all available pages before exploring further. # Limits and errors > How MCP tool calls consume your plan, what happens when a limit is hit mid-conversation, and how to read the errors your AI client shows. MCP runs on the **same entitlements as the REST API**. There is no separate MCP allowance, and nothing here is charged differently: a tool call is a request. ## How a conversation consumes your plan | | | | ---------------- | ------------------------------------------------------------ | | One tool call | One request against your monthly allowance | | One question | Often **several** tool calls | | A JSON-RPC batch | Charged **per `tools/call` inside it**, not once per request | | Per-minute burst | Shared with REST, same ceiling | | Inference | Not billed by us. Your AI client uses your own subscription | A single question routinely fires three to ten tool calls: the model may list expiries, fetch levels, then pull flow before answering. Budget by questions rather than by requests, and assume a handful of requests per question. See [Rate limits](/concepts/rate-limits) for the per-plan numbers. Because a burst of calls arrives inside one answer, the starter burst allowance is set well above one call per question. Hitting the ceiling mid-answer is the failure this protects against. ### Batching Most clients send one tool call per request and never think about this. If yours batches, two rules apply: * Each `tools/call` in the batch is charged separately, so a batch costs the same as sending its calls one at a time. * **A single batch may carry at most 50 `tools/call` messages.** Beyond that the whole batch is rejected with a JSON-RPC `Invalid Request` reading `too many tools/call in one batch; the limit is 50 per request`, and nothing in it runs. Split it and resend. Fifty is far above what a question needs (three to ten is typical), so this is a guard against runaway automation rather than a limit normal use meets. ## When something goes wrong Errors are returned to your AI client as readable text, so you generally see an explanation in the conversation rather than a status code. ### You are not connected The client shows an authorization prompt, or reports that the server requires sign-in. Reconnect from your client's settings. Tokens are tied to the BackQuant account that approved them, so this also appears after you revoke a connection. If you connect with an API key rather than OAuth, the same situation reads as `the access token is invalid or expired`. That means the key was mistyped, revoked, or belongs to a plan without API access. Check it at the API dashboard rather than reconnecting, since there is no sign-in step to repeat. ### Your plan does not include API access You are told during sign-in, before the connection completes, rather than by a confusing tool failure later. Terminal Monthly has no API access; a trial does not grant it either. See [the plan table](/mcp-server) for what qualifies. ### Your subscription is not active Same point in the flow as above. Access resumes when the subscription is active again; nothing needs reconnecting on our side. ### Monthly quota exhausted The tool returns the API's own message, which names the cap, and your client relays it mid-conversation. Nothing is broken; the allowance is spent for the calendar month. Quotas run per user, not per key, and reset at the start of the next month in UTC. ### Too many calls too quickly A burst above the per-minute ceiling is rejected until the window rolls. In practice this means waiting a few seconds and asking again. ### The data is not available Some answers do not exist for some markets: a symbol with no aggressor tape, an expiry that is not listed, a history window longer than what is stored. The tool says so rather than returning an empty result that reads as zero. ### A tool failed unexpectedly You will see the tool name and the error type. Nothing internal is exposed to the conversation. If it repeats, send us the tool name and roughly when it happened. ## Response size Every tool has a token budget, and responses are shaped to stay inside it. A raw options chain would be correct and useless: it would fill your client's context window and leave no room for reasoning. So tools return ranked, trimmed answers rather than complete dumps. Practically, that means asking for "the top strikes by gamma" works well, and asking for "every strike and every expiry" gets you a sensible subset. If you need the complete surface, use the [REST API](/api/v2/gex/strike-profile), which has no such budget. ## Coverage boundaries * **Crypto only.** Equities are not available through the API or MCP. * **BTC and ETH** have the fullest coverage, including multi-venue aggressor tape. SOL and HYPE are supported where data exists. * **Historical depth varies by dataset** and is shorter than live coverage. If a window comes back shorter than you asked for, that is the stored depth, not an error.