Rate limits
Per-minute and concurrent-run limits by plan, the headers that report them, and how to back off from a 429.
Rate limits apply per API key over a sliding one-minute window. Starter and Pro allow 60 requests a minute and Enterprise allows 300. Every plan gets the same headers.
Response headers
Every response, successful or rejected, carries:
| Header | Meaning |
|---|---|
X-RateLimit-Limit | Allowed requests in the current window. |
X-RateLimit-Remaining | Requests left before the window resets. |
X-RateLimit-Reset | Unix timestamp at which the window resets. |
Retry-After | Seconds until the next allowed call. Only set on 429. |
Reading those is enough to back off without polling.
These headers report the rate limit, a short sliding window. The
separate X-Quota-* headers report the monthly plan budget; see
Billing.
What rejection looks like
HTTP/1.1 429 Too Many Requests
X-RateLimit-Limit: 60
X-RateLimit-Remaining: 0
X-RateLimit-Reset: 1747396800
Retry-After: 12
Content-Type: application/json
{"detail": {"code": "rate_limited", "message": "Per-key rate limit exceeded."}}Plan budgets
| Plan | Requests per minute, per key |
|---|---|
| Starter | 60 |
| Pro | 60 |
| Enterprise | 300 (negotiable per contract) |
Pricing lists every plan limit in one table.
Concurrent runs
Each account can have a limited number of agent runs in flight at once, across all of its keys and MCP:
| Plan | Runs at once |
|---|---|
| Starter | 2 |
| Pro | 5 |
| Enterprise | 10 |
One more returns 429 too_many_concurrent_runs with Retry-After.
It isn't billed; retry once a run finishes.
MCP endpoint limits
The MCP endpoint also limits requests per client IP, before any tool
runs. Every MCP message counts, including initialize and tools/list. IPv6
clients are counted per /64 network.
| Requests | Per minute, per IP |
|---|---|
| With a well-formed API key | 1,500 |
| Without a key, or with a malformed one | 30 |
Invalid API key on ask_neuroapi | 20 |
Over a limit, the HTTP response is 429 rate_limited with Retry-After, and
X-RateLimit-* on MCP responses describe these IP limits. After too many invalid
keys, API and dashboard sign-in requests from that IP also get 429 until the
window passes, even with valid credentials. Your plan's per-minute budget still
applies to each ask_neuroapi call.
Surge handling
Retry a rate_limited 429 after Retry-After. A quota_exceeded 429 has no Retry-After and only clears at X-Quota-Reset, so this sample doesn't retry it:
Always cap retries - a 429 storm with unbounded recursion will exhaust your stack. The example below makes up to three retries.
async function ask(body: unknown, attempt = 0): Promise<Response> {
const res = await fetch("https://api.neurobro.ai/api/v1/agent/ask", {
method: "POST",
headers: {
"Content-Type": "application/json",
"X-API-Key": process.env.NEUROAPI_KEY!,
},
body: JSON.stringify(body),
})
const retryAfter = res.headers.get("retry-after")
// A quota_exceeded 429 has no Retry-After, so it is not retried.
if (res.status === 429 && retryAfter !== null && attempt < 3) {
const wait = Number(retryAfter)
await new Promise((r) => setTimeout(r, wait * 1000))
return ask(body, attempt + 1)
}
return res
}For long-running batch jobs, prefer pacing yourself with
X-RateLimit-Remaining so you never see 429 at all.
Capacity errors are different
503 Service Unavailable from /agent/ask means the agent is saturated
right now. It does not mean you exceeded your quota, and no quota was
charged. Retry shortly with exponential backoff.