Now Telepath sits in ChatGPT too

Telepath spent three weeks going through OpenAI’s review.

Today it was published in the ChatGPT directory.

Before that it had already passed Anthropic’s review, appeared in the Claude Directory and shipped separately as a Claude Plugin.

Now the same Telegram can be connected to Claude and to ChatGPT through a single layer — Lemma Telepath.

For a project that was a folder of code on my server not very long ago, that is a rather odd point to have reached.

This is the place where one wants to write:

Everything is ready now.

But the past few weeks have broken me of that sort of phrasing fairly thoroughly.

So it is more accurate to put it this way:

Telepath now stands, for the first time, in front of people I cannot explain anything to in person.

It all started with one question

The first version of Telepath was not a product at all.

I wanted to be able to write to a model:

Have a look at what people wrote to me on Telegram today.

And for it not to explain why it has no access to Telegram, but genuinely go there and look.

What came out was a simple chain:

Telegram → archive → MCP → model

Then came search, reading conversations, files, photographs, voice messages, reactions and the ability to reply wherever Telegram permits it.

Then the same MCP connected to Claude.

And then a more interesting question turned up: if I can use this, why can somebody else not use it the same way?

That is the point where the project stopped being merely a server, quite quickly.

Along came:

OAuth.
Separate users.
Archive isolation.
Permissions.
A privacy policy.
Terms.
Public infrastructure.
Subscriptions.
History import.
Onboarding for a person who has never heard the word MCP.

And the project acquired a name — Lemma Telepath.

The diagram now looks different:

Telegram → Telepath → ChatGPT / Claude

Telegram is connected once.

After that, different assistants get the permitted tools for working with one and the same personal archive.

An important part: this is not a Telegram emulation

Telepath works through the official Telegram Business API.

No MTProto.
No logging into a user’s Telegram session.
No handing the service an account password or a session file.

The user adds @lemma_telepath_bot as a Telegram Business bot themselves, and decides themselves which chats it can reach.

And one more important detail: Telegram Premium is not required to connect Telepath.

On current clients the path looks like this:

Telegram → Settings → My Account → Chat automation

That is where @lemma_telepath_bot is added.

On older clients it may live under Telegram Business → Chatbots.

Once connected, new messages start arriving in the personal Telepath archive.

Chats the user has excluded from the bot’s access never reach it at all.

And older history can, if needed, be imported separately from Telegram Desktop.

So Telepath does not get “access to all of Telegram”.

It gets exactly the stream that Telegram officially hands to a connected Business bot.

That boundary matters to me.

Claude came first

In early October Telepath passed Anthropic’s review and appeared in the Claude Directory.

For me that was already a noticeable boundary.

Until then the MCP existed as a technical entity.

You could take the server URL.
Connect it by hand.
Read the README.
Work out what a remote MCP is.

A catalogue changes the very way a product shows up.

What a person sees first is not the protocol.

They see:

Lemma Telepath
Your Telegram, in Claude

And somewhere far below that sit OAuth, SQLite, webhooks, the Business API and dozens of MCP tools.

It seems like a small difference until you walk that path yourself.

A developer almost always builds bottom-up.

A user sees the system, for the first time, top-down.

Later Telepath appeared as a Claude Plugin as well.

Which means that inside Claude’s ecosystem it is no longer just a link to an MCP server.

The plugin holds instructions and separate skills for the typical scenarios: connect Telegram, work quickly through a pile of unread conversations, prepare a brief on a discussion or a draft reply.

One and the same backend has gradually started turning into a product across several surfaces.

Then the OpenAI review began

When I first submitted Telepath to OpenAI, I expected questions about architecture.

OAuth.
Message storage.
Authentication.
The Telegram API.
User isolation.

But the first problems turned out to be far more interesting.

What they reviewed first were the tool descriptions.

If an MCP tool promises the model a little more than the Telegram API actually permits, that is already a problem.

If an annotation creates the wrong impression about where a request goes, that is a problem too.

A public tool must not only work correctly.

It must describe the boundaries of its own work to the model with absolute precision.

We corrected the wording.
Rebuilt the submission.
Clarified the permissions.
Updated the product description.
Submitted again.

In total, roughly three weeks passed between the first submission and publication.

And those three weeks turned out to be a separate phase of development.

The MCP itself was working by then.

But between “the server works” and “it can be placed inside ChatGPT in front of a stranger” there turned out to be a whole further layer of product work.

16 October

Today the process ended with the word I had deliberately avoided writing early for three weeks:

Published.

Lemma Telepath has appeared in the ChatGPT directory.

The path this all started for now looks almost literal.

You open ChatGPT.
You find Telepath.
You connect it.

And after that you can ask:

What did people write to me on Telegram today?

The model does not receive text I pasted into the ChatGPT window beforehand.

It calls Telepath’s tools and works with the real archive.

It can look at recent conversations.
Find a person.
Search across the whole correspondence.
Read a file.
Look at a photograph.
Transcribe a voice message.

And, where the rights allow, reply or set a reaction.

In September you had to understand what an MCP was and where to paste a URL.

On 16 October Telepath became an installable object inside ChatGPT itself.

For me that is the main change.

The server did not learn anything fundamentally new overnight.

One more layer of knowledge, previously required of the user, simply disappeared.

It has made for an odd mirror

The same Telegram can now be opened from two very different assistants.

Claude knows Telepath through its own ecosystem.

ChatGPT through its own.

For the user that is two interfaces, two catalogues and two products around a model.

From underneath, Telepath barely sees a difference.

An authorized client arrives.
It has a user.
The user has a separate archive.
The model calls the permitted tools.

What I like about this construction is precisely its symmetry.

Not:

I built an integration with ChatGPT

and separately:

I built an integration with Claude

But rather:

I built a layer between Telegram and AI assistants.

Today the client is called Claude.

Another is called ChatGPT.

Tomorrow it might be called something else.

Telegram should not have to change because of that.

And nor, as far as possible, should Telepath.

The most interesting part begins after the review

Until today, nearly every problem in Telepath was found by people and systems that were deliberately trying to test it.

Me.
Tests.
Scanners.
Anthropic’s reviewer.
OpenAI’s reviewer.

Those are rather comfortable users.

They either understand how the system is built, or at least know that it needs careful checking.

A person from the catalogue is under no obligation to do any of that.

They should not have to know the difference between Telegram Login and a Telegram Business bot.
Should not have to understand MCP.
Should not have to read documentation.
Should not have to guess which setting to open.

And they certainly owe the product no second chance if the first attempt turned out to be incomprehensible.

They will simply press Install.

And from there the system either carries them to a working state by itself, or it does not.

So publication right now does not feel like the end of the testing.

Rather the opposite.

Until now I was checking whether Telepath talks correctly to Telegram, to MCP clients and to models.

Now it will become visible for the first time whether it talks correctly to a person who first saw the name Telepath ten seconds ago.

Where Telepath is now

As of 16 October the picture is this.

Telepath works through the official Telegram Business API.
Premium is not required to connect.
Every user has a separate archive.
There is OAuth and a public hosted service.
There is search across conversations.
There are files and photographs.
There are voice messages and transcription.
There are replies, edits and reactions wherever Telegram grants the corresponding rights.
There is an import of older history from Telegram Desktop.
There is group support.
There is the official MCP Registry.
There is the Claude Directory.
There is the Claude Plugin.

And as of today there is a public listing in ChatGPT.

It is still the same project that began with a fairly everyday thought:

It would be convenient simply to ask the model what people have written to me.

Except that now I am not the only one who can ask that question.

And today, it seems, is when that became genuinely testable.