tiny.help
Pricing Open desk Start free trial

Solving tickets with ChatGPT and Codex

One URL puts your support inbox inside ChatGPT. It reads the thread, drafts the reply in your voice, and waits for you to say send.

Start your free 14-day trial See how to connect

Works with the ChatGPT plan you already have. Developer mode, one MCP URL, and nothing to host.

78 seconds, nothing cut: switch on, URL in, signed in, and ChatGPT reading the inbox.

The whole setup, start to finish

Four steps and a sign-in. You do this once, on one browser, and it’s on your whole ChatGPT account.

  1. Developer mode on

    Settings → Plugins → Developer mode. It’s the switch that lets ChatGPT talk to a connector OpenAI didn’t build.

  2. One URL

    New plugin, name it, paste desk.tiny.help/mcp, leave auth on OAuth. Nothing to host, no key to copy.

  3. Sign in as you

    The same Google account you use for your desk. Whoever presses Allow is who ChatGPT acts as.

  4. Ask it something

    “Can you check conversations in tinyhelp”. Seven open threads, newest first, desk never opened.

Codex answers with the code in front of it

A visitor reports a bug with a screenshot. get_conversation brings the report, the pages they were on and when, their browser and OS, and the screenshot itself into your terminal.

Codex is already sitting in your source with your logs a command away, so it goes from “the button does nothing” to the line that broke it, fixes it, and drafts the reply. The customer hears back before you’ve switched windows.

codex

Answer from wherever ChatGPT is

The connector is added once in a browser and then it’s on the account: the phone app, the desktop app, chatgpt.com. Ask what came in, read a thread, get a draft in your voice, say send. tiny.help doesn’t even need to be open.

Set your voice notes once and every draft reads like you wrote it, whichever device you’re holding.

ChatGPT app

The tools. One rule.

tiny.help is a remote MCP server. ChatGPT and Codex both already know how to use one; there’s nothing to install on your side beyond the connection.

  • list_conversationsYour inbox, newest first, with a snippet of the last message. Filter by site, by open or closed, or by one customer’s name or email.
  • get_conversationOne full thread with the quoted-email pyramids already stripped, the visitor’s browser, OS, location and recent pageviews, and attachment URLs it can open to look at screenshots.
  • get_voice_notesYour tone, stock phrases, sign-offs and the things you never say. The server tells the model to read this before drafting anything as you.
  • send_replySends an already-approved reply as you. The customer sees it in their chat widget immediately; if they’ve gone, the email pipeline follows up.
  • react_to_messagePuts a ❤️ 👍 👎 😂 or ❓ on one message, the way a tapback works in Messages. The customer sees it in their chat widget right away, and it can be taken back.
  • close_conversationWrap it up. A new message from the customer reopens it on its own.
  • list_shortcutsThe saved replies on each of your sites: the stock answers your team expands in the composer by typing /pricing.
  • save_shortcutWrite a new saved reply, or rewrite one you already have. It hands back the wording it replaced, so an overwrite is never a one-way door.
  • delete_shortcutRetire one. It comes back with the body it removed, in case you want it again.

And the setup and knowledge tools, for the day you connect: say “Set up tiny.help for https://yoursite.com” and it walks you through the whole thing.

  • setup_status, setup_guideWhere the account stands, what is left to do, and the script to follow.
  • start_checkoutThe Stripe link for a 14-day trial. You add the card; it comes back.
  • inspect_siteReads your homepage: brand colours, sitemap shape, sign-in links, whether the widget is already there.
  • create_site, update_site, get_site, list_themesThe site, its welcome message, colours, quick links, live monitoring and the embed snippet.
  • update_knowledge, start_crawl, list_pages, delete_pagesWhat to index and what to skip, the crawl, and pruning what got in.
  • get_faq_guide, upload_faq_filesA prompt for writing your support guides from your code or your docs, shown to you in full first, and the upload once you have read every word of the result.
  • get_corrections, set_correctionsThe standing notes the assistant reads before every answer. “Prices went up on 1 September.”
  • invite_member, set_member_sites, list_teamYour team, inside the plan's limits.
  • set_assistantSwitches the assistant on for a site, once there is something for it to answer from. It is off until you ask for it.

You approve every send. That’s not a marketing line; it’s the instruction the server ships with the send tool, every time it’s listed:

“This delivers to a real person and cannot be unsent. Never use this tool to compose, preview, or ‘show’ a reply. Write drafts in the chat as plain text instead. ‘Reply to #3’, ‘draft one’, silence, or your own suggestion to send are none of them approval. If you are unsure whether you have it, you do not. Draft and ask instead.”
Keep send_reply on ask, never on always allow. Customer messages are untrusted text, and your approval of each send is what stops one of them speaking as you. The assistant is the other thing, and it is deliberately not the same: it answers on its own, but only from pages you approved, only in four categories, and it goes quiet the second one of your people types. The one with your voice never sends without you; the one that talks to customers can only ever quote you.

Connect once

ChatGPT and Codex are two separate setups: a connector on your account, and a line in your Codex config. Both use the same sign-in and the same access rules as the desk itself.

ChatGPT: web, desktop and phone

Added once in a browser, it’s on your whole ChatGPT account.

  1. In ChatGPT, open Settings → Plugins and scroll to Developer mode. Turn it on.
  2. Back on Plugins, press + for a new plugin. Name it TinyHelp.
  3. Paste the URL above as the server URL, leave Authentication on OAuth.
  4. Tick the box to say you understand the risk, then Create → Connect.
  5. Sign in with the Google account you use for your desk. Whoever presses Allow is who ChatGPT acts as.

On your phone it turns up in a new chat under + → the plugin list, on the same account, without setting anything up again.

Codex CLI

One computer, two commands. Adds nothing to your ChatGPT account or phone.

codex mcp add tinyhelp \
  --url https://desk.tiny.help/mcp
codex mcp login tinyhelp
the sign-in opens in your browser

The same sign-in as the desk. From then on Codex can work a ticket while it’s sitting in your project’s code: read the bug report, find the line, draft the reply.

Questions

Do I need a paid ChatGPT plan?

Custom MCP connectors live behind Developer mode, which is a paid-plan feature, and the account in the video is a Plus one. Without it you still get the whole desk, and the assistant that answers your customers still works: it runs on our account and is included on every plan. What a paid plan buys you is the setup done for you, and the ability to work a thread without opening the desk.

Why do I have to turn on Developer mode?

It’s the switch that lets ChatGPT talk to an MCP server OpenAI didn’t build and hasn’t reviewed. tiny.help is one of those. It’s a per-account setting under Settings → Plugins → Developer mode, and you only flip it once.

Why does ChatGPT warn me about custom connectors?

Because it should. A custom MCP server can read what you hand it and act on your behalf, and OpenAI hasn’t reviewed ours. That warning is generic to every custom connector, not a finding about this one. Read what the tools do before you tick the box. All of ours are listed above, and the setup tools on the Claude page work the same way from Codex or ChatGPT. None of them touch anything but your own desk.

Does ChatGPT ever send a reply on its own?

The send tool only accepts a reply you’ve already approved, and its own description tells the model that a draft, “reply to #3”, or silence isn’t approval. On your side, leave send_reply on ask rather than always allow: that confirmation is the last thing standing between a customer’s message and your name on a reply.

Does this work in Codex?

Yes, and it’s one command: codex mcp add tinyhelp --url https://desk.tiny.help/mcp, then codex mcp login tinyhelp. Codex is running in your repository, so it can check the code the customer is asking about before it drafts anything.

Does tiny.help read my source code?

No. tiny.help hands over the conversation: the messages, the visitor’s browser and pages, any screenshots. Codex reads your code because it’s running in your repository. That’s what makes the answer right.

Can my teammates use it too?

Everyone on your desk connects with their own sign-in. Drafts come out in their voice notes, and a send is theirs. The customer sees who replied.

What about Claude?

Same server, same tools, same rule. The Claude page has its own video and its own two-minute setup.

What does the connector cost?

It’s included on every plan. $29 a month is three sites and three people, with a 14-day trial. The drafting runs on your ChatGPT plan.

Try it on your own inbox

Install the widget, paste one URL into ChatGPT, and answer the next hard question from wherever you are. 14 days free, card up front, nothing charged until day 15.

Start your free 14-day trial