What is Gephyr?

Gephyr makes your app usable by AI assistants like ChatGPT, Claude, Grok, Meta Muse and Cursor, without writing code.

You connect your existing API, choose what assistants may do, and Gephyr publishes a secure connector for your app. Your users add it to their assistant, sign in with the account they already have, and then just ask: “Log a 30-minute run”, “What classes are on Thursday?”, “Move my booking to 6pm”.

How it works

Your user ──asks──▶ AI assistant ──calls──▶ Gephyr connector ──calls──▶ Your API
                                    (mcp.gephyr.ai/t/yourapp) (as that user)
  1. You connect your API. Gephyr turns your endpoints into a small set of clear actions that assistants use well.
  2. You choose what’s allowed. Each action is read, write, confirm (preview plus approval) or off.
  3. Users sign in with their existing account. They log in on your own login page or, for apps on Firebase or Supabase Auth, on a sign-in page with your branding that sends the password straight to Firebase or Supabase. The password never reaches Gephyr.
  4. Gephyr publishes and hosts it. You get a connector link, a branded setup page for your users, and templates to announce it.
  5. You see what happens. Every call is logged, along with what users wish your app could do.

What Gephyr handles for you

  • Tool design: task-level actions, not a raw list of endpoints. Long lists come a page at a time, and answers are kept to a size agents can read (see what agents get back). Apps with more than 40 actions get a search, so agents find the right one among hundreds. See apps with many actions.
  • Sign-in: works with Auth0, Okta, Clerk, Amazon Cognito, Supabase, Firebase and any OAuth/OpenID provider.
  • Safety: access levels, single-use approvals for high-risk actions (on the assistant’s own approval screen where it has one), a prompt injection guard (text your company didn’t write, like reviews and messages, is labelled so assistants show it rather than follow it; Confirm still guards what matters), and strict limits on what Gephyr will call.
  • Hosting: a fast global endpoint that works in every assistant that supports MCP, the open standard for connecting assistants to apps. It speaks the latest version (2026-07-28) and the earlier ones assistants still use, so nobody has to update anything.
  • Distribution: a setup page, announcement templates, a public server card, an optional listing in the official MCP Registry (under Gephyr’s name or your own domain), and help getting listed in assistant directories. See where people find it.
  • Paying customers only: give your users a reason to upgrade by keeping your connector, or some of its actions, for users on your own paid plans. See who can use it.
  • Message board (beta): share news, service notices and offers that your users’ assistants pass on when they ask. See message board.
  • Gephyr AI (every plan): what to fix next, worked out from how agents use your app, with clearer descriptions, notes and new actions written for you when you ask. See Gephyr AI.
  • Test with AI (paid plans): see whether AI models pick the right action for what your users ask, before you publish, with clearer descriptions suggested for any they mix up. See test your actions.
  • Keeping up with your API: Gephyr reads your API docs every day and tells you which actions a change would break, before your users find out. See when your API changes.
  • Privacy: your users are a pseudonymous ID to Gephyr, never a name or email, and calls are logged without what was asked or answered. See privacy at Gephyr.
  • A CLI for you and your AI agent: set up, import, choose levels and publish from a terminal, or paste one prompt into Claude Code, Cursor or Codex and let it set your app up from your codebase. See Gephyr CLI.

Where to go next

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