Agentkit's Rate Limit setting allows a set number of incoming visitor messages from one visitor key, usually an IP address, during a time window. The default is 20 requests in 240 seconds. Once that key has used its allowance, further requests in the window get your limit message instead of an AI answer. This can slow a noisy visitor, but it does not cap monthly spending across a workspace. Set the value from actual conversations, then test what a visitor sees when it runs out.
What the setting counts
Open your chatbot's Settings → Security → Rate Limit. The control accepts 1–100 messages, a window of 10–3,600 seconds, and a limit message of up to 500 characters. The on-screen controls read "Limit to [number] messages every [number] seconds" and "Message to show when limit is hit." The rate-limiting guide explains the fields.
Each chatbot has its own allowance. The first request after the previous window has expired starts a new fixed window. Requests within that window count toward the maximum; a blocked request does not extend its end. Requests can bunch up around a reset, so do not treat the limit as a steady requests-per-minute ceiling.
The settings screen calls this a limit from "one device." For website chat, the server actually uses the client IP when available and falls back to the conversation ID. Without an IP, starting a new conversation can start a fresh allowance. That key is not a person. Two devices on the same office network may share the allowance; changing networks may produce a fresh one. The setting cannot cap the number of different visitor keys. If the limit check itself errors on the server, the request is let through.
Workspace credits count a different event: one saved AI reply uses one credit. A visitor's incoming message alone does not. Images and video can spend from the same shared balance. Check the usage and credits guide for the balance and reset rules. A blocked request returns the configured text without saving an AI reply, so it does not spend a chat credit. The sidebar balance is shared by every chatbot and other generation in the workspace, so do not use its change alone to diagnose the limiter.
Choose a limit from real conversations
Review Chat Logs for legitimate conversations that took several turns. Count requests within the window you plan to use, especially for the longest useful support or buying questions. This is a starting point: Chat Logs show conversations, while the limit can group several conversations behind one IP. Leave room for shared traffic if many visitors use one office or campus network. If questions repeat because an answer is missing, fix the source as well.
Write a limit message that explains the pause and gives an existing contact route if the matter cannot wait. For example: "This chat has reached its message limit. Please try again later or use the support link on this page." Do not promise a ticket, human handoff or exact reset time unless you have checked that behavior on your site.
Test the blocked message and recovery
Use a test chatbot on a page you control, with enough credits for the test. Keep other people and other chat surfaces, including the Playground, off that chatbot during the test. Requests from the same network can change your count.
- Set the maximum to 3 and the window to 120 seconds. Save the distinctive message "Test limit reached." If you already sent messages to this chatbot from the same network, wait 120 seconds after the last one before starting.
- From one device on one network, without switching Wi-Fi or using a VPN, note the time of the first website-widget request and send four short messages within about 90 seconds. The fourth should return exactly "Test limit reached." instead of an AI answer. This limit text appears in the widget but is not saved as an AI reply. A request that fails after the limit check can still count; if one of the first three fails, wait for a fresh window and restart the test.
- Wait about 130 seconds from your first counted request before sending one more. It should get an AI answer in a new window. The blocked fourth request did not restart the clock.
- Open that conversation in Chat Logs. The expected result is saved AI replies for messages 1–3 and 5, with no AI reply for message 4. The blocked fourth response in the widget and an answered fifth request are the primary pass conditions.
- Set the intended values on the live chatbot and confirm they saved. Test a normal customer question sequence there to check that it is not interrupted. These requests use the tester's allowance and spend credits when AI replies are saved.
This test checks the website widget on one network. If the fourth request is not blocked, reopen Security settings to confirm the values saved, confirm the test page embeds that same chatbot, then check whether the network changed, the four messages took longer than 120 seconds, or a request failed before reaching the limiter. If the failure persists after a quiet-window retest, contact support; a visitor response cannot reveal a server-side limit-check error. If a request is blocked early or recovery is late, another request from the same network may have used the allowance. Restart after a full quiet window. Repeat the test after changing the setting.



