Yes, in production. The server is at https://api.tbit.app/mcp with Streamable HTTP transport. With ChatGPT and Claude you connect with your TBit account: you add the URL, tap Connect and log in. There is also a manual token for advanced clients.

How do I connect ChatGPT to TBit?

In ChatGPT open Settings, go to Connectors and add a connector with the URL https://api.tbit.app/mcp. Tap Connect: the TBit login opens, you log in and authorize the permissions. No tokens to copy. Custom connectors are available on ChatGPT paid plans.

Do I need to copy a token?

No: with ChatGPT and Claude all you need is the URL and to tap Connect and log in. The manual token exists only for advanced clients, like scripts or MCP clients without login support.

What can ChatGPT do when connected to my agent?

With read-only access it can see the prompt, memories, recent conversations and unanswered questions. With the training scope it can create, edit and delete memories and save a prompt draft that you activate in the app.

Is it safe to give my assistant access?

Yes. You authorize with your account and see the permissions before connecting. Access can be revoked with one click, with a limit of 60 requests per minute, and it covers your agents according to your role: you only train where you are an administrator. The assistant never sees the internal configuration of the TBit engine and any prompt change goes through your approval.

Help center

TBit MCP: connect ChatGPT or Claude to your AI agent

TBit has an MCP server to connect ChatGPT or Claude to your agent and train it with your data: memories, conversations and prompt. URL, token and how to do it.

Published on July 18, 2026Updated on October 10, 20265 min read

What it is

TBit exposes an MCP server (Model Context Protocol) so your AI assistant (ChatGPT, Claude or any MCP client) can connect to your TBit agent and help you train it with your real data: read its configuration, review memories and conversations, create memories and suggest prompt improvements. Your assistant acts as a trainer; sensitive changes stay as a draft that you approve in the app.

You no longer need to copy tokens: with ChatGPT and Claude the connection is made with your TBit account, you log in and authorize the permissions. For advanced clients the manual token is still available.

Connect from ChatGPT or Claude (recommended)

ChatGPT (Plus or Pro): Plugins, Add, Create custom MCP server. Paste the URL, leave authentication on OAuth, check "I understand and want to continue" and tap Create as plugin. You no longer need to turn on developer mode; in ChatGPT for business the admin turns it on.
Claude: your account menu, Settings, Connectors, Add, Add custom connector. Paste the URL, tap Continue and Add without changing authentication, then Connect.
In both cases the TBit login opens: you sign in and tap Allow. Done, no tokens to copy.

What you'll see when you authorize

You'll see the permissions "View your agents: prompt, memories, conversations and guide" and, if your assistant asks for it, "Also create and edit memories and suggest prompt improvements (you approve them)". The connection covers your agents based on your role in each one: view with any role, train only where you're an admin. If you have more agents tomorrow, the connection covers them on its own.

Disconnect

You can disconnect from your assistant (remove the connector) or wait for it to expire. Access lasts 1 hour and renews itself while the connector is active.

Connection details

Manual token (advanced)

For clients without login connection support (scripts, some MCP clients), you can use a manual token:

Go to Agent Settings, Connect your AI section.
Generate a token with the scope you need: read only or training.
Paste it as authentication in your MCP client.

The token is per agent. It's shown only once, so save it well, and you can revoke it whenever you want from Settings; access is cut off instantly.

What your connected assistant can do

These are the tools your assistant sees when it connects, based on the permissions you authorize.

If you have several agents, your assistant will ask you which one to work with, or you can tell it by name.

Read only

list_my_agents: lists your agents with their role.
get_agent_overview: the agent's name, business and prompt.
list_memories and search_memories: the agent's memories, with search by meaning.
get_recent_conversations: recent conversations.
get_training_backlog: questions the agent couldn't answer.
list_guide_topics and get_guide_topic: the platform guide.

Training

create_memory, update_memory and delete_memory: create, edit and delete the agent's memories.
save_prompt_draft: suggests the prompt as a draft; you activate it in the app.

Starter prompt

Paste this into your assistant after connecting the connector:

Eres el entrenador de mi agente de TBit. Conéctate, llama get_agent_overview y list_guide_topics, y ayúdame a mejorar el entrenamiento: revisa el prompt, las memorias y las conversaciones recientes antes de proponer cambios.

Security

You authorize it with your TBit account, and the permissions are shown before connecting.
The connection covers your agents according to your role: view with any role, train only where you are an administrator.
You can disconnect whenever you want; access lasts 1 hour and renews itself while the connector is active.
The assistant never sees the internal configuration of the TBit engine.
The assistant suggests and you approve: prompt changes go through your approval in the app before taking effect.