Gephyr AI
Gephyr AI helps you build your connector and keep it working well. When you import your API, it turns your docs into a short list of clear actions. Once you’re live, Gephyr watches how AI agents use your app and points out what to fix: actions agents call wrongly, answers that get cut short, actions that keep failing, fields that look private, things your users ask for that your API can already do. Most of these are worked out from numbers and your API docs. Where a fix needs words, like a clearer description, Gephyr AI can write it for you.
Gephyr AI only ever helps you design and improve your connector. It never acts for your users: their own AI agent does that, with the actions you’ve chosen.
For the whole picture, what Gephyr handles automatically and what helps in your API docs, see Make your API agent-ready.
Nothing changes until you say so. A suggestion becomes an unsaved change you can read, edit or drop, and nothing reaches AI agents until you publish.
When you import your API
Every import is designed by Gephyr AI, automatically, on every plan. It doesn’t count toward your allowance. It:
- Picks the actions a typical user would want (up to 20) and leaves out admin, internal and deprecated endpoints, telling you what it left out and why.
- Names and describes each action in plain words, and keeps any rule from your docs that agents need (“ask for at least one of these”, “dates must be in the future”). If it leaves a rule out, Gephyr adds the sentence back from your docs.
- Describes inputs in plain words an agent can act on: what goes there, in what form, with a short example where it helps. It rewrites descriptions that are missing, unclear or written for developers, using only what your docs show, and never changes an input’s name, type, limits or whether it’s required. Your docs’ example values and rules are kept, and a rule the input itself sets (numbers only, a date) is said in words.
- Suggests access levels: it can make an action more careful than Gephyr’s rules, never less.
- Writes the example requests for your setup page, like “Show my runs from last week”. Suggest examples on Publish & share writes a fresh set.
You review all of it before anything is saved or published. See From endpoints to actions and Make your API agent-ready.
Two kinds of suggestion after launch
- Gephyr suggests comes from rules: your app’s call counts, your actions and your API docs. It’s always on, on every plan, and doesn’t count toward anything.
- Gephyr AI text is written by an AI model and carries the Gephyr AI badge. You ask for it with a button, and each answer uses one from your workspace’s monthly allowance.
Where you’ll see it
| Where | What Gephyr suggests | What Gephyr AI can write |
|---|---|---|
| Overview, “What to fix next” | Up to six things, worst first: an action failing in the last 24 hours, agents sending inputs an action refuses, answers cut to fit, feature requests your API could answer, fields that look private | An explanation of why an action is failing and what to try |
| Actions, editing an action | Fields that look private and aren’t hidden; a paging setup found in your API docs, for lists that get cut short | A clearer description, written from the inputs agents get wrong. Rules your current description gives agents are always kept |
| Actions, How agents use your app | A draft of Notes for AI agents from your actions | |
| Activity, feature requests | Requests matched to endpoints in your API docs, each as a ready action to add, at your API’s address | |
| Protection | A higher daily limit when people keep reaching it, or a lower one when even your busiest user is far below it |
When “What to fix next” has nothing to show, your app is in good shape.
When a suggestion appears
- Inputs agents get wrong: at least 5 calls in the last 7 days, and at least 1 in 10 refused for bad input.
- Answers cut to fit: at least 5 successful calls in the last 7 days, at least 1 in 5 cut, and the action doesn’t page yet. If your API docs show how the list pages, you can turn paging on in one click. If the action has its own page inputs (like
limitorcursor), turning paging on takes them out, since Gephyr’s paging gives agents its own. See What agents get back. - A failing action: at least 5 calls in the last 24 hours and at least 1 in 20 failing.
- Feature requests: at least 3 in the last 30 days, and endpoints in your API docs that no action uses. Matching them needs your docs link (set it under Keep up with your API on the Actions tab; see When your API changes).
- A price the approval may not show: an Ask first action whose answer, in your API docs, names an amount (like
totaloramount_charged) with no price check. If one of its inputs already holds the price, you can turn the check on in one click; if not, it explains how to add one your API checks. - Private-looking fields: names like
email,phone,address,date_of_birth,card_numberorapi_keythat an action returns and doesn’t hide. Fields you’ve marked as written by other people aren’t flagged. When the fields are inside a list your API wraps (like{ "data": [ … ], "has_more": true }), they’re named from that list, and hiding them also sets List starts at to it, so the agent gets the list itself.
Feature requests and private fields come from your API docs, which Gephyr reads at most once an hour for this.
Your monthly allowance
Import and example requests are free, with their own monthly allowance of AI drafting per workspace: 20 with no plan yet, 30 on Starter, 50 on Launch, 100 on Growth, 200 on Scale and 500 on Enterprise. Each import from your API docs, and each set of example requests, uses one when the AI writes it; a draft the AI couldn’t finish doesn’t count. Once it’s spent, imports and examples carry on with Gephyr’s rules (one action per useful endpoint, up to the first 60, and simple examples) until the 1st of the next month. Imports and example requests are also limited to 10 a minute from one address.
The suggestions you ask for with a Gephyr AI button after launch have a monthly allowance per workspace (UTC calendar month), shared by all its apps:
| Plan | Gephyr AI suggestions a month |
|---|---|
| No plan yet | 10 |
| Starter | 15 |
| Launch | 20 |
| Growth | 60 |
| Scale | 150 |
| Enterprise | 500 |
A suggestion only counts when Gephyr AI comes back with something. If the model can’t be reached, or it has nothing better than what you have, it doesn’t count. The count shows next to each Gephyr AI button, and under your plan at the bottom of the sidebar. It starts again at 0 on the 1st of each month (UTC); unused suggestions don’t carry over. If you change plan, the new plan’s allowance applies straight away. Gephyr suggests never counts.
What the model sees
Gephyr AI sees your app’s name and description, your actions (names, descriptions, inputs and levels), your notes for agents, your API docs, and counts such as “10 of 40 calls refused for bad input”. To match feature requests, it also sees the requests’ text: the words users chose to send, which you can already read on Activity, never who sent them.
On import, it sees your API docs and the draft actions Gephyr built from them.
It never sees your users’ identities, what they asked agents in chat, or what your API answered. Gephyr doesn’t call your API for any of this. Suggestions are written through Gephyr’s AI providers. See Privacy.
Something missing or wrong? Email hello@gephyr.ai.