Protection

Gephyr stops runaway or abusive AI agents before they reach your API. Every app gets sensible limits from the start, and you can change them on the Protection tab of your app in the console.

Calls Gephyr stops never reach your API and are never billed. The agent gets a plain message saying what happened and what to do (wait a minute, or that today’s limit is reached), and the stops show on your app’s Activity page, under Why calls failed. Overview shows how many people hit a limit, and the Protection tab shows the same for the last 7 days.

Limits for each user

How often one person’s AI agent can use each action. Each action is counted on its own, so a user who hits the limit for one action can still use the others.

• Per-minute limits come from a fixed set of choices: 10, 20, 30, 60, 120, 300, 600, 1,200 or 3,000, or no limit.
• Per-day limits can be any whole number up to 1,000,000, or no limit. They reset at midnight UTC.
• For Confirm actions, each preview counts toward the day. Approving it doesn’t use up another call from the day, but approvals have their own per-minute limit, so a turned-down approval can’t be retried over and over.
• A “user” is the person signed in through your login.
• Not sure what to set? When people keep reaching a daily limit, or even your busiest user is far below it, Gephyr suggests a better one under the table, with one click to use it. Save to keep it.
• Apps without login: Gephyr can’t tell users apart, and AI apps like ChatGPT and Claude call from their own servers, so many people can share one network address. Limits then apply per address at ten times the per-minute limit (60 becomes 600), with no daily limit. The app-wide ceiling still applies.

Limits for single actions

Give one action a tighter or looser limit than the rest, for example 20 bookings a day, or no limit on a look-up your users call often. Everything else keeps the limits above.

Whole app

• All users together: a ceiling for sudden floods across everyone, 1,200 calls a minute by default. Only calls within each user’s own limits count toward it, so one user going over their limits can’t use it up for everyone else. It’s counted separately in each region Gephyr runs in, so it’s approximate: a flood spread around the world can go a little over.
• Stop repeated changes (on by default): when an agent sends exactly the same Write action with the same inputs twice within 10 seconds, only the first reaches your API. This stops double bookings and double posts from agents that retry. The agent is told to check whether the first one worked, and can call again after 10 seconds if it didn’t or the user really wants it done twice.
• Approving in the AI app, for Confirm actions, is set with the other agent settings, under How agents use your app on the Actions tab. See how Confirm works.

Feature requests

Each user’s agent can send up to 10 feature requests a minute. The same request sent again within a few seconds is acknowledged but recorded once. Requests over the limit aren’t passed on; each shows on your dashboard as a stopped request_capability call, like other limit stops.

Good to know

• Each user’s per-minute limits, feature-request limits and the repeat guard are counted exactly, over a rolling minute (or 10 seconds). Only the whole-app ceiling is approximate. Daily limits are counted exactly too, though a call or two can get through while the stop takes effect. Raising a daily limit lets users who reached the old one carry on as soon as it’s published.
• Saved changes go live when you publish, like other settings. Use Gephyr’s defaults goes back to the limits above, and follows them if they change.
• Limits protect your API and your plan’s calls. Your own API’s rate limits still apply.

Something missing or wrong? Email hello@gephyr.ai.