Telegram in Claude, ChatGPT, Grok and AI agents

Telepath

Work gets agreed in a messenger, and the assistant cannot see any of it. Telepath opens that correspondence to its owner’s assistant — over the official Business API, without handing over a user session.

year
2026
context
Our own product
role
Architecture and storage, the Telegram collector, our own OAuth server, text extraction from attachments, transcription, permissions and subscriptions, landing page, documentation, catalogue submissions.
live
telepath.lemma.companyClaude directoryproduct channel
reading
how to connect, step by step

the problem

Correspondence is working memory: the agreements, the files and the voice notes that never made it into any system. An assistant sees mail, tasks and code — and precisely not this.

The ready-made ways to give it access ask for a user session, which is the key to the whole account. For working correspondence that is not an option, and offering it to other people even less so.

  1. telegram
  2. account archive
  3. MCP
  4. assistant
a message reaches the assistant with nobody forwarding it

what the thing had to be

  • the official way

    the Telegram Business API and a bot the person adds to their own account

    which also means that API’s limits — other people’s chats stay out of reach

  • separation

    one archive per account, physically its own

  • more than reading

    attachments, photos and voice — otherwise the assistant sees fragments of a conversation

  • any client

    one archive and a standard MCP endpoint instead of an integration per vendor

how it works

The person adds the bot to their Telegram Business and picks which chats it sees. From then on new messages arrive by webhook, land in that account’s archive and become available to whichever assistant the owner authorised.

Attachments do not stay as file identifiers: text is pulled out of documents and spreadsheets, photos are handed over as images, voice notes as transcripts. Otherwise a search finds the word “contract” but not what the contract says.

  1. bot in the account Business API
  2. webhook new messages
  3. account archive one per person
  4. text and voice attachments, transcripts
  5. MCP search, read, reply
  6. assistant client of your choice
from a message in a chat to an answer in the assistant

what the assistant gets

  • search

    full-text across every conversation, captions and transcripts included

  • a thread

    the history of one chat, not just the last few messages

  • attachments

    PDFs, spreadsheets and documents by their content, not their file name

  • voice

    transcripts of voice notes and round videos

  • replying

    sending, editing and reactions as the owner, if they switched it on

  • the past

    import of a Telegram Desktop export: the bot only sees what came after it was connected

signing in

Authorisation turned out to be a project of its own. The client registers itself, by the standard, with no keys handed out in advance; the person proves who they are with Telegram; the token that comes back belongs to that account and is checked against its audience on every request.

identity
Telegram Login — the messenger vouches for the person, there is no password
protocol
OAuth 2.1 with PKCE and dynamic client registration
binding
a token works for one account and is checked against its audience
scope
a client granted read-only never sees the tools that send

one service, several platforms

The archive is not tied to one assistant: the owner picks the client, the correspondence stays the same.

  • Claude

    published in the connector directory

  • ChatGPT

    submitted to the catalogue, in review

  • Grok

    connects as a custom connector; a PR is open to the catalogue of their terminal agent

  • agents and CLI

    the standard remote MCP endpoint — no separate integration needed

said plainly

A whole product: from somebody else’s API and a storage schema to sign-in, a landing page, documentation and a line in a major platform’s catalogue.

There is no end-to-end encryption here, and we do not pretend otherwise: to search and transcribe, the server processes messages, which means it technically can read them. What is done about it: archives are separated per account, the server’s disk is encrypted whole, and a command in the bot erases everything we hold. Anyone who needs a guarantee that no operator will ever read a line runs the same code on their own machine, single-tenant.

ask