NewAssistant Labs for ChatGPT

Run your whole business from ChatGPT

Connect your Assistant Labs account to ChatGPT once. Then just ask — what came in this week, who is still waiting, why the agent got an answer wrong. It reads your real conversations and fixes the agent while you talk to it. In Codex, the same account runs the day for you and brings you the decisions.

Connects in about five minutes
Sign in — no API key to paste Works in ChatGPT and in Codex Nothing reaches a customer without your yes
ChatGPT Assistant Labs
Which customers are still waiting for an answer?
Usedlist_conversations

Four people are still waiting. The longest has been 19 hours:

Yael — asked whether you deliver to Eilat. The agent never answered.

Dana — asked for a quote on 40 units.

We do deliver to Eilat. Teach the agent that.
Usedpatch_agent_module
Added to the delivery FAQ. Try asking it now.
Built on the Model Context Protocol
ChatGPT on the webChatGPT desktopCodex CLICodex in your IDE

What actually changes

The work stops being yours

Not a chatbot with tools bolted on. The difference is who does the work and who makes the call.

Today
  • You read every message to find the three that matter
  • You remember which customer is still waiting
  • You write the same reply for the fifth time this week
  • You notice the lead went cold a fortnight too late
  • Nothing happens while you are with a customer, driving, or asleep
With ChatGPT connected
  • It reads everything and tells you the three that matter
  • Nobody waits, because it knows exactly who is owed a reply
  • The reply is drafted in your voice before you ask
  • A lead going quiet becomes a chase on a timer, not a regret
  • It works overnight and hands you the decisions in the morning

Three connectors

Your whole operation, in one chat

Each connector opens up a different part of Assistant Labs. Add the ones you use — they all work side by side in the same conversation.

Agents

Conversations, customers and the agent itself

Every conversation your agent has had, the contacts behind them, and the settings that decide how it answers.

  • Read, search and summarise real conversations
  • Edit the agent — FAQs, catalog, hours, tone
  • Contacts, labels and segments
  • Send WhatsApp templates and replies
Connector addresshttps://mcp-server-150134556021.us-central1.run.app/mcp

Sales

Journeys, audiences and outbound

The WhatsApp journeys that follow up on carts, orders and quiet customers — and how they are performing.

  • Review and edit journeys
  • Check who is enrolled and who converted
  • Manage message templates
Connector addresshttps://mcp-server-150134556021.us-central1.run.app/sales/mcp

Tasks

The work your business owes people

What is open, what is blocked, and what someone promised a customer — kept in one place instead of in your head.

  • Open, update and close tasks
  • Ask what is still outstanding
  • Turn a conversation into a task
Connector addresshttps://mcp-server-150134556021.us-central1.run.app/tasks/mcp

Every connector is a read of your live account — nothing is copied, exported or stored on ChatGPT’s side.

It reads the real thing

Your actual conversations — not a summary of them

ChatGPT reads the transcripts your customers actually sent: the questions, the follow-ups, and the moments the agent lost them. Ask in plain language and it finds the pattern across hundreds of chats.

  • Search by what was said, not by keyword tags
  • Volumes, busiest hours, repeat questions
  • Full transcripts, including Hebrew
  • Long answers are capped and always say when they are partial
What ChatGPT uses
  • list_conversations
  • search_conversations
  • conversation_stats
  • get_conversation

It can act, not just report

Fix the agent in the same breath you spot the problem

Finding a bad answer and fixing it are usually two different tools and two different days. Here they are one sentence. ChatGPT edits the agent’s knowledge, and you can have it rehearse the new answer before anything goes live.

  • Edit FAQs, catalog, hours, links and tone
  • Add instructions in your own words
  • Re-scan your website when something changes
  • Simulate a conversation before you save
What ChatGPT uses
  • patch_agent_module
  • set_extra_instructions
  • scan_website
  • simulate_conversation

It reaches your customers

From “who’s gone quiet?” to a message sent

The same conversation that surfaces a segment can write to it. ChatGPT pulls the contacts, drafts the message in your voice, and sends it through the channels you have already connected.

  • Contacts, labels and saved segments
  • Create and send WhatsApp templates
  • Reply into a live conversation
  • Every send obeys the permissions you approved
What ChatGPT uses
  • list_contacts
  • list_segments
  • create_whatsapp_template
  • send_whatsapp_template

Things to ask it

No dashboards. Just ask.

These are ordinary sentences, typed into ChatGPT. It works out which part of your account to read and answers from your real data.

Understand the week

How many conversations came in this week, and what did people ask about most?

Show me every chat where someone asked about delivery and didn’t get a straight answer.

Which customers are still waiting for a reply?

What are people complaining about lately?

Fix the agent

The agent keeps saying we’re open on Saturday. We’re not — fix it.

Add a returns policy to the FAQ, in Hebrew.

Re-scan our website; we changed our prices.

Before you save that — show me how the agent would answer.

Reach customers

Who bought from us twice but hasn’t been back in three months?

Draft a WhatsApp update for everyone labelled VIP.

Reply to Dana and tell her the quote is on the way.

Or name a skill

In Codex the operator installs as skills. Ask in your own words, or name the one you want — these are the ones you will want most days.

  • startBegin here. It works out where you are and takes you one step forward.
  • briefWhat is happening right now — who is waiting, what is at risk, today’s three things.
  • needs-meEverything waiting on your decision, one at a time, ready to approve.
  • waitingEvery customer who has not been answered — with the replies already drafted.
  • leadsWho to chase, warmest first, with what to say to each.
  • moneyWhat is owed, what is at risk, what is coming in.
  • agent-reviewWhere your agent answered badly this week — and the fix.
  • autopilotA full unattended pass. Does everything. Sends nothing.

The deal

Three tiers, and they never move

The whole thing rests on this. It is not a setting you tune — it is the contract.

It just does it

Reading, thinking, preparing

Everything reversible that nobody outside the business ever sees.

  • Reads and analyses every conversation
  • Drafts replies, reports and campaigns
  • Keeps your customer records straight
  • Files what needs doing, so nothing is lost
It does it, then tells you

Small fixes you could undo in a minute

It acts, keeps the receipt, and mentions it once.

  • Corrects a stale price your agent was quoting
  • Fixes a customer’s phone number
  • Re-files a task, adds a label
It stops and asks

Anything a person or a shekel touches

Every time. An approval yesterday is not an approval today.

  • Every message that reaches a customer
  • Every charge, refund, discount or price change
  • Anything that cannot be undone
  • Anything that changes what someone is charged

It will never send a message because nobody was around to say no. An owner who is away is not a quiet yes — the decision waits for them, with the finished text attached.

When you are not at a screen

It reaches you on WhatsApp. Tapping Approve is the last step.

From your own business number — the one your customers already message and you already have saved. Not a second sender to learn to trust.

  1. 1

    It asks, with the finished thing attached

    Not "shall I draft something for Dana?" — the exact message she would receive, word for word, so there is nothing left to do but decide.

  2. 2

    You tap, or you type

    Approve, No, or just write back — "yes", "not yet", "send it". It understands Hebrew and English, and it never reads a "don’t" as a yes.

  3. 3

    It happens, then it tells you what happened

    Not "approved" — the actual outcome. If a send failed you hear that instead, because being told it went out when it did not is the one thing this design exists to prevent.

Replying never talks to your bot. Your own agent never answers you, never files you as a lead in your own customer list, and never sees the conversation.

It learns your business

It sounds like you, because it read how you write

A generic assistant is worthless. What makes it yours is a set of pages it keeps in your workspace and reads before it acts.

Your voice, from your real replies

Not "friendly and professional" — actual sentences your people sent to actual customers. Drafts come back sounding like the business.

Your rules, in your words

Tell it once that you do not deliver on Saturdays and it lands in the right place with the date, and fixes your agent if that was wrong too.

Who does what

If Dana handles invoices and decides refunds up to ₪500, a ₪380 refund is Dana’s call and never reaches you at all.

Kept with your account rather than on one computer — so it still knows you from a new laptop, and when it works overnight with nobody there.

Security

It only ever gets what you hand it

A connector is a door into your business, so it is built to be narrow, inspectable and easy to shut.

You approve each permission

The consent screen lists exactly what ChatGPT is asking for — reading conversations, editing the agent, sending messages. Untick anything you would rather keep to yourself.

Read-only stays read-only

Tools are built from the permissions you actually granted. A read-only connection is never even shown a tool that could change or send something.

Nothing is stored on the way

The connector holds no copy of your data and no credentials of its own. Every question becomes one live request to your account, answered and forgotten.

Sign in — don’t paste a secret

You connect with your Assistant Labs login, so no key ever lands in a URL, a screenshot or a chat log. Access is bound to that one connector and expires on its own.

One org, one agent

A connection reaches only the organization you signed in to, and only the agents inside it. If you run several, you pick which one it may see.

Revoke in one click

Disconnect it in ChatGPT, or cut it off from your Assistant Labs settings. Access stops immediately — there is nothing left behind to clean up.

And the things it never does

Your permissions are enforced by the server, not asked of the model — a connection that cannot send is never handed a tool that sends.

  • Sends, spends or deletes without a specific yes
  • Stores a password, an API key or a card number
  • Acts on instructions it read in a customer’s message
  • Approves its own actions
  • Quietly stops doing something that was working
  • Tells you it is done instead of telling you the customer got their answer

Setup

Two ways in. Start with either.

Connect it to ChatGPT and you can ask your business anything. Install the operator in Codex and it runs the day for you. Most people start with ChatGPT and add Codex later.

A

Connect it to ChatGPT · about five minutes

Custom connectors live in ChatGPT’s developer mode. Nothing is downloaded and nothing runs on your computer — you paste one address and sign in.

1

Turn on developer mode

In ChatGPT, open Settings → Apps & Connectors → Advanced and switch on developer mode. On a Business, Enterprise or Edu workspace an admin has to allow it first.

2

Add the connector and sign in

Choose Create, paste the connector address from Assistant Labs, and let ChatGPT scan the tools. Sign in with your Assistant Labs account and approve what it may do.

3

Publish it to yourself, then ask

A new app starts as a draft — publish it to your own workspace, open a chat and ask it about your week.

B

Install the operator · Codex

The same account, one level up: it runs the day unattended and brings you the decisions. Codex reads one config file and a folder of skills — no marketplace, no developer mode.

  1. 1

    Add the connectors to Codex

    Put the Assistant Labs servers in your Codex config, once per computer.

    open ~/.codex/config.toml
  2. 2

    Authorise each one

    Signs you in the same way ChatGPT does, in your own account. Each connector is its own decision.

    codex mcp login assistantlabs
  3. 3

    Install the operator

    The skills that make Codex run the business rather than just read it.

    cp -R skills/* ~/.codex/skills/
  4. 4

    Ask for what you want

    Say "brief me", or "who is waiting". Codex loads the matching skill on its own — there is nothing to memorise.

Several steps are optional — say no and setup is still finished

  • The number it reaches you on
  • The WhatsApp templates that carry approvals
  • Who else works with you, and what they decide
  • A scheduled morning brief
Works everywhere OpenAI does
ChatGPT in the browser The ChatGPT desktop app Codex in the terminal Codex in your IDE

Questions

The things people ask first

Still unsure about something? Message us and we’ll answer properly.

Ask us on WhatsApp
Do I need a paid ChatGPT plan?
Custom connectors need developer mode, which is available on ChatGPT’s paid plans. On a Business, Enterprise or Edu workspace an admin has to allow it first. Codex has no such gate — any account can add the connectors from a config file. On the Assistant Labs side there is nothing extra to buy.
Is this an app I have to install?
In ChatGPT it is a connector: you paste one web address and sign in. Nothing is downloaded and nothing runs on your computer. In Codex it is a config file and a folder of skills you copy in.
Can it change things without asking me?
It can only do what you approved when you connected. Connect with read-only permissions and it cannot change or send anything at all — that gate is in our server, so it holds whichever app you are using.
Is this as locked down as the Claude version?
The permissions are identical, because they are enforced by our server rather than by the app. One difference is worth knowing: the Claude Code plugin adds a local check that catches an outgoing message before it is sent, and OpenAI’s apps have no equivalent hook for connector tools. The operator still stops and asks in both. If you want the hard stop, connect read-only and approve sends yourself.
What is the difference between ChatGPT and Codex here?
ChatGPT is for asking — your conversations, your customers, your agent, in a normal chat. Codex is where the operator lives: it runs the day, drafts the replies and brings you the decisions. Same account, same permissions, and you can use both.
Does it work in Hebrew?
Yes. Your conversations are read as they were written, and you can ask your questions and get your answers in Hebrew.
What if I have more than one agent?
You choose which one the connection may reach when you approve it. If you leave several available, just tell it which agent you mean.
Can I use Claude instead?
Yes — the same connectors work with Claude, and there is a Claude Code plugin as well. There is a page for that too.
How much does it cost?
Connecting is included with your account, and the operator is free. Both run on your Assistant Labs workspace, which is free to start.

Your business already answers questions all day. Now you can ask it some.

Connect Assistant Labs to ChatGPT, find out what your customers have been asking while you were busy — and let Codex start handing you the decisions instead of the work.

Start free Free to start · No card · Nothing sends without your yes