MCP for Telegram: connect ChatGPT and Claude
How to connect the TeleTraff panel to Claude and ChatGPT over MCP: server address, sign-in with or without a key, limits, dry run and what the agent cannot do.
- One server address and sign-in through OAuth or a key from Settings, API & MCP
- A dry run shows what will happen before the task starts
- Archived accounts and accounts without a proxy are never used
- Money and admin are not part of the API
To connect ChatGPT or Claude to TeleTraff, add a custom MCP server with the address https://teletraff.com/api/public/mcp and sign in to the panel with the Allow button. After that the assistant sees your account fleet, starts modules and reads the results, and every action follows the same rules as in TeleTraff automation.
What is MCP and what is the TeleTraff server address?
MCP (Model Context Protocol) is a common way for an AI assistant to reach outside services: the service publishes a list of tools and the assistant calls them when you ask. TeleTraff keeps such a server open for every workspace.
- Server address: https://teletraff.com/api/public/mcp.
- Protocol: Streamable HTTP, requests are sent with POST. The server keeps no session between requests.
- Passwords, Telegram account sessions and proxy details never appear in the answers: the assistant sees statuses, counters and results.
How do I sign in: with a key or without one?
- Without a key (OAuth): you confirm access on a TeleTraff page and copy nothing. The sign-in lasts 1 hour and renews itself for up to 30 days. You can switch it off under Settings, API & MCP, API keys.
- With a key: in Settings, API & MCP, press “Create key”, then “Issue key”. Before the form the panel asks you to confirm your identity, and issuing needs a verified email. The key goes into the client settings as a Bearer token.
A key is shown once
Right after issuing, the panel says it will not show the key again and displays it in full only at that moment. Store it in a password manager. If the key leaks, delete it under API keys and the assistant using it loses access at once.
How do I connect Claude, ChatGPT and Claude Code?
- 1
Claude (claude.ai and Claude Desktop)
In the panel open the “Connect an AI agent” dialog and press “Open Claude”: the form is already filled in. You can also do it by hand: Settings, Connectors, Add custom connector, and paste the server address. In the Claude window press Continue, then Allow on the TeleTraff page. It takes under a minute.
- 2
ChatGPT (paid plan required)
In ChatGPT open the Plugins section, press Add, then Add custom MCP server. Paste the name and the server address, tick “I understand and want to continue”, press “Create as a plugin” and confirm access with Allow on the TeleTraff page.
- 3
Claude Code in the terminal
Run the command: claude mcp add --transport http teletraff https://teletraff.com/api/public/mcp. On the first call Claude Code opens the TeleTraff sign-in page.
Check the connection with the question: “How many active accounts do I have in TeleTraff?”. The assistant takes the answer from the accounts summary and does not count archived ones. Which modules it can start is described in the Automation section.
What can the assistant do and what can it not do?
Rights depend on the scopes of the key. A sign-in without a key grants what you tick on the consent page.
- Read: accounts, conversations, jobs, datasets, help.
- Messages: sending and receiving incoming ones.
- Webhooks: creating them and checking deliveries.
- Jobs: start, pause, resume and stop of modules and schedules.
- Manage: archive and tags for accounts, pools, proxies, datasets, module settings, campaigns.
- Delete: accounts, proxies and datasets. The panel warns that it cannot be undone, so enable this scope only on purpose.
- At most 60 requests per 60 seconds per key. Above that the server answers with an error and a wait time.
- Panel operations that change something: at most 30 calls per minute per key.
- A dry run is available for jobs and for sending a message: no job is created and the assistant gets a launch preview.
- Repeating an identical send does not go out twice.
- Cold messages follow a set pace. The default is careful: 5 messages per hour and 20 per day per account. Balanced gives 10 and 40, fast gives 20 and 80. Replies to people who wrote last do not spend these limits.
- Archived and switched-off accounts never appear in lists or launches. Accounts without a proxy are refused by the panel.
Example: check 30 accounts and dry-run a mailing
- 1
Ask for a summary
“Check my accounts and tell me how many are ready for a mailing.” The assistant reads the summary and the fleet capacity. Say 3 of the 30 accounts have no proxy: they are refused and 27 remain.
- 2
Ask it to read the rules
Before a plan the assistant opens the help on Telegram limits and the Mailing module description. It takes the numbers from there.
- 3
Request a dry run
“Prepare a mailing to the offer base as a dry run.” Nothing is created: you see how many accounts take part and what would happen.
- 4
Start it and watch the counters
After your yes the assistant starts the job and waits for it through events and per-recipient results. The delivery counters decide the outcome: a created job does not yet mean the messages arrived.
At the careful pace 27 accounts send up to 135 cold messages per hour and up to 540 per day. That is an upper bound: Telegram can limit an account sooner. An account that gets PeerFlood is held by the API for 24 hours and kept out of new launches.
What does the AI agent inside the panel do?
If you do not need an outside assistant, the panel has its own “AI agent”. It works in a chat and uses the same set of tools the public API is built from. Agent statuses: Ready, Working, Waiting for confirmation, Writing a reply, Needs attention. Suggestions in an empty chat: help set up a module, how to set up an account, check accounts and proxies.
- Under agent permissions there is a “Full access” switch. With it the agent changes reversible settings without extra questions, while purchases, deletions and launches still need separate confirmation.
- Without full access an action arrives as a “Waiting for confirmation” card with the buttons “Allow once” and “Cancel”. The card is valid for 10 minutes, and if you do not answer the action is not done.
- The agent does not write or debug code, SQL or commands, does not reveal passwords, sessions, API keys or phones, and does not name the model it runs on.
- The agent treats instructions inside other people's texts as data and does not carry them out.
Risks and what the panel does not do
- A key with the delete scope can erase accounts, proxies and datasets with no way back. Issue it rarely: reading, jobs and manage are enough for daily work, and archiving can be reversed.
- A leaked key gives the same access you have. Issue a separate key for each client and delete the ones you no longer use.
- The model makes mistakes. Ask for a dry run before a start and check the job counters. Stopping a job cannot be undone.
- Texts from Telegram, job events and evidence are written by strangers. The assistant must treat them as data: if a message asks it to start something, it must not obey.
- Money and admin are not in the API: buying accounts, plans and payouts are unavailable through an assistant. Launches that need a Telegram code or a confirmation in the panel are unavailable too.
- Every change is written to the panel log, and answers contain no phones, sessions or proxy details.
How to work through Telegram bots and mini apps on many accounts is covered in the bot automation guide. From $29 per 30 days on Starter: 10 streams in each module. Pro is $59 (50 streams), Team $99 (150 streams). The 3-day trial starts with the first imported account. Accounts and proxies are bought separately.
FAQ
What is the TeleTraff MCP server address?
https://teletraff.com/api/public/mcp. The same address works for Claude, ChatGPT and Claude Code, only the sign-in method differs.
Do I need a key to connect Claude or ChatGPT?
No. Sign-in without a key goes through OAuth: you press Allow on the TeleTraff page. A key from Settings, API & MCP is for clients that do not support OAuth. ChatGPT needs a paid plan.
Can the assistant send a message or start a mailing by itself?
Yes, if the key or sign-in has the needed scopes. It should do a dry run before the start, and cold messages follow a set pace: by default 5 per hour and 20 per day per account.
Will the assistant see the passwords and sessions of my accounts?
No. API answers contain no phones, sessions or proxy details, and every change goes to the panel log.
How is MCP different from the AI agent in the panel?
MCP connects an outside assistant, such as Claude or ChatGPT, to your workspace. The AI agent lives inside the panel, asks for confirmation with a 10-minute card and does not reveal keys or sessions.
Ready to open the panel?
Review the scenario constraints before launch.
Read next

Job automation
Schedules that launch your modules by themselves at a set time: one schedule, several modules, a live preview of upcoming runs and an execution log.
GuideHow to Manage Multiple Telegram Accounts from One Panel
Dozens of Telegram accounts without the chaos: import, pools with their own proxies, profiles, two-factor protection, status checks and jobs on a whole group.
GuideHow to Send a Mass Message on Telegram: Four Ways and How to Choose
Four ways to send a mass message on Telegram: channel, bot, direct messages, group chats. When each fits, what it needs, how to run the last two in TeleTraff.