Getting started

Most apps go from sign-up to a published connector in 10 to 15 minutes. You don’t write any code.

Setting up is four steps, the same as on the homepage and in the console: your app, what agents can do, login and publish. Then you watch it work.

Availability: importing takes OpenAPI, Swagger and Postman files in JSON or YAML; HAR recordings are coming.

1. Show us your app

Sign up at gephyr.ai and choose how to start. Add your app’s name and website; logo and brand colour can come later. You immediately get your connector’s addresses (the link can’t be changed later):

Then upload an OpenAPI or Postman file, or paste a link to it. Gephyr’s AI proposes a short list of actions in plain language. See Connect your API.

Working in Claude Code, Cursor or Codex? Choose Let your AI agent do it instead: you paste one prompt into your coding agent and it does this step and the next one from your codebase with the Gephyr CLI, checking with you as it goes.

2. Choose what agents can do

For each action, choose Read, Write, Confirm (shown as Ask first in the console) or Off, and adjust names and descriptions. See Tools and access.

3. Connect your login

If your actions work with a user’s own account, set up how users sign in. Enter your domain: Gephyr detects your auth provider and, for most providers, sets everything up with one Connect button. See Sign-in.

Then run Test sign-in with your own account. Gephyr checks three things: who you are, a real API call, and a token refresh. You can publish once all three pass. See Test sign-in.

Public-data connectors (no user accounts) skip this step.

Once your workspace is on Launch or up, right after Gephyr’s AI picks your actions, Test with AI runs realistic requests on Claude and GPT to see whether they pick the right action, and suggests fixes for any mix-ups. See Test your actions.

Once agents start using your app, Gephyr AI lists what to fix next on Overview (actions agents get wrong, answers cut short, fields that look private) and can write clearer descriptions, notes for agents and new actions for what your users ask for. Nothing changes until you accept a suggestion.

4. Publish and tell your users

Publish. The first time, Gephyr asks you to choose a plan: plans start with Starter, $9.99 a month for 10,000 calls (everything before this step is free; see Plans and billing). Then link your users to your setup page and use the announcement templates. To be found without your link, switch on the MCP Registry listing under Where people find it. See Publish and share.

Then: watch it work

Your app’s Overview shows:

• calls, active users, success rate and response time, with the change since the same time in the previous period;
• calls per day (or per hour for the last 24 hours), in your time zone: hover over or tap a bar for its details;
• on Growth and up, where calls come from: which AI apps (with each one’s people, success rate, speed and main actions; with Ask first actions, also whether that AI app shows its own approval screen or the preview in the chat) and which countries the requests came from; hover over or tap a row for details;
• on Growth and up, how each action is doing (typical and slowest response times, failures);
• people: first-time and returning users, new connections and disconnections, how many people hit one of your limits, and (on Growth and up) which agent apps people connected through;
• the path from setup page visits to connecting to a first call;
• on Growth and up, your busiest days and hours, in your time zone (hover over or tap a square for the count).

If an action starts failing, a note at the top says so.

Notices. Each note at the top of Overview (an action failing, approved actions to check, API docs that changed, or a safety warning about your live app) has a close button (×). A closed note stays closed until what it’s about changes, such as another failed call, another approved action to check or a different action failing; then it comes back. The Notices bell in the sidebar counts the open ones across all your apps and lists them, each with a link to where to fix it, and you can close them there too. New notices pop up once in the corner when you open the console. What you close is remembered for your account, so it stays closed on every computer you sign in on, and each teammate closes their own.

Check these, at the top of Overview on every plan, lists ask-first actions a user approved whose outcome is unclear: Gephyr sent them, then your API timed out, had a server error or sent an answer Gephyr couldn’t read. Gephyr never repeats these, so look each one up in your own system by its time, action and AI app (no user details are shown). If one didn’t happen, the user can simply ask again.

Activity explains why calls failed with what to do about each kind (including approved actions with an unclear outcome, and answers too large for an AI agent) and shows how users answered ask-first actions, on every plan. It lists recent calls (and failed ones on their own) on Launch and up, and groups the things users asked for that your app can’t do yet on Growth and up. On other plans those places show how many there were. Users show only as anonymous IDs.

Gephyr can see which agent app made a call (Claude, ChatGPT, Cursor…) but not which AI model it used, or what the user typed.

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