What your AI client sees when it's close to a limit.

Burst capacity is generous (≈ 100 seconds of full-throttle calls) so well-behaved AI clients rarely see cap_exceeded. If yours is hitting a limit, the cause is almost always parallel fan-out without coordination — serialise calls or reduce concurrency.

import time

MAX_RETRIES = 3
CAP_EXCEEDED = "cap_exceeded"

def call_with_backoff(call_tool, name: str, arguments: dict):
    """call_tool is your MCP client's tool-call function.
    JSON-RPC errors surface differently per SDK (return value
    or exception); normalise to (code, text, result) first."""
    for attempt in range(MAX_RETRIES):
        code, text, result = call_tool(name, arguments)
        if code is None:
            return result

        if code == -32601:
            # Tool is not on your tier. Retrying never helps.
            raise PermissionError(f"{name} is not available on this tier")

        if code == -32000 and CAP_EXCEEDED in text:
            # Rate limiter, concurrency cap, or a daily quota.
            # A daily quota will not clear until the window resets,
            # so keep retries bounded.
            time.sleep(2 ** attempt)
            continue

        raise RuntimeError(f"Gateway error {code}: {text}")

    raise RuntimeError(f"{CAP_EXCEEDED} after {MAX_RETRIES} retries")

If your AI client hits cap_exceeded frequently, the right move is usually to reduce parallel fan-out (run searches sequentially within a workflow rather than firing all jurisdictions at once). If the sustained workload genuinely needs more capacity than your tier provides, talk to us about the Company tier.

If your workload is bumping against the current limits, get in touch — Company-tier capacity is contracted, and we'd rather size it with you than have you engineer around a cap.