POST /boot and POST /chat to keep one visitor or one workspace from overwhelming the service. Limits apply at two levels: per visitor and per workspace. When a request crosses a limit, the worker rejects it with 429 and tells you when to retry.
The 429 response
A throttled request returns status429 with a JSON body and a Retry-After header (in seconds):
Retry-After header and wait at least that long before retrying. Some responses also carry the same value as a retryAfterSec field.
Burst windows report a fixed 60-second
Retry-After. Sustained per-visitor windows report the seconds remaining in the current window.Limits
A chat request can be checked against the visitor burst, workspace burst, and sustained per-visitor windows in turn. Requests that arrive without a valid visitor token fall back to the per-IP limit.
Monthly plan allowances
Throttling protects the service from bursts. Separately, each plan includes a monthly number of agent actions. One agent action is one change the guide completes for a visitor, such as a confirmed click or a form fill. Pointing, highlighting, scrolling, navigation, and generated answers never count.
Every plan can make confirmed writes up to its allowance. A spent allowance never returns
429 or 402: chat keeps answering and the guide points at a control instead of acting on it. Growth, Scale and Business can add extra actions when overage is enabled with a monthly cap.
When the month’s AI replies are spent, chat still never returns 429 or 402, and the widget never goes blank. A conversation already under way can finish with up to three more answers. After that the guide writes no new answer: it shows the visitor the most relevant pages of your knowledge as sources. When none fit, it hands the question to your team if human handoff is on, or shows the contact you set under Guides → Actions → Contact for visitors, or says the assistant can’t answer right now. The visitor is never told about your plan. On a paid plan, usage packs in Billing keep it answering, up to the cap you choose.
Writes to a target whose Consequence is financial, contractual, or irreversible, including unclassified controls Puntego detects as one, need Growth or higher. See Verified Actions.
Back off cleanly
1
Honor Retry-After
On a
429, wait for the Retry-After interval before the next attempt. Do not retry immediately.2
Add exponential backoff with jitter
If retries keep failing, grow the delay on each attempt (for example 1s, 2s, 4s) and add a small random offset so concurrent clients do not retry in lockstep.
3
Cap retries and surface the limit
Stop after a few attempts rather than looping. The embed runtime already paces its own requests, so persistent
429s usually mean traffic is genuinely above the limit.Next steps
Errors
Every error namespace and status code, including the rate-limit envelope.
Authentication
How visitor sessions are minted and how programmatic API keys are scoped.
API reference
Request shapes and responses for the worker endpoints.