---
title: Conversations
description: Talk to assistants in Conversations: pick one, attach a file, approve the actions it wants to run, and find past conversations with search and pins.
url: https://agentlab.cresclab.com/docs/en/studio/conversations
language: en
updated: 2026-11-06
---

# Conversations

Talk to assistants in Conversations: pick one, attach a file, approve the actions it wants to run, and find past conversations with search and pins.

## Start a conversation

Click **Conversations** in the console's left rail. With no conversation open, the right side of the page is a message box: type, send, and a new conversation starts.

1. **Check which assistant gets it** Under the box it says **Sends to … Change assistant**. Click **Change assistant**, or type `@` and an assistant's name in the message.
2. **Type the message** Up to 4,000 characters. Type `/` and a skill's name to put that skill's text into your message; you see what was added before sending.
3. **Click Send** The conversation appears in the list on the left. Its name is generated for you, and you can change it later.

A conversation holds at most 100 messages. When it is full you see **This conversation is full. Start a new one to keep going.**

## Choose an assistant

You can talk to two kinds of assistants: shared ones published to the workspace, and your own private ones. A draft has to be published first.

- While the workspace's **General assistant** switch is on, a new message goes to the general assistant unless you pick another. The switch is in **AI settings** and is managed by owners and admins.
- With it off, the only assistant is chosen for you; with several, the console asks you to **Choose an assistant**. An assistant missing from that list can still be found by typing `@` in the message.
- If there is nothing to talk to yet, the page offers **Go to assistants**.

A conversation stays with one assistant. To switch, start a new conversation.

Each message can also use its own **Model** and **Thinking effort**. Without a choice it **Uses the model set on the assistant**. A model your plan doesn't include is refused.

To create and publish assistants, see [Assistants](https://agentlab.cresclab.com/docs/en/studio/assistants.md).

## Attach a file

Click **Attach a file** in the message box to have the assistant read a document.

| Item      | Limit                                         |
| --------- | --------------------------------------------- |
| File type | Text (`.txt`), Markdown (`.md`), PDF          |
| Size      | Text or Markdown up to 2 MiB, PDF up to 5 MiB |
| Count     | One file per message                          |

A PDF needs selectable text; a scanned image PDF can't be read. If a file is longer than the model can take in, shorten it or choose another model.

Files belong to the workspace and stay after the person who uploaded them leaves. Open **Conversation details** from the menu at the top right of a conversation to export a file under **File export and retention**. Exports are recorded in the audit trail.

## Approve what the assistant wants to do

Some tool actions need a person's go-ahead first. The conversation then shows a card, **The assistant wants to run an action**, listing what it will send.

| Label on the card                            | Who can approve            |
| -------------------------------------------- | -------------------------- |
| **Read-only**                                | You                        |
| **Writes to an external system**             | A workspace owner or admin |
| **Undeclared risk — treated as destructive** | A workspace owner or admin |

**Approve and run** runs the action and the assistant carries on with its reply. **Decline** has it answer without the action. Declining needs no admin rights.

On a card that needs an owner or admin, a member sees only **Decline**, with the note **Only a workspace owner or admin can approve this action. Ask one of them to approve it here, or decline it.** The card waits until an owner or admin opens the conversation and approves it, or you decline. Owners and admins see **Approve and run** on the same card, without the note. Once approved, the action still runs as you, with your account.

Other things you may see:

- **Search the web** and **Read a web page**: the query or address goes to the public web only after you approve.
- **Ran automatically (this workspace lets this tool run without approval)** means the workspace lets this tool run without anyone approving it.
- **Not run — a rule in this workspace does not allow it.** A workspace rule stopped the action. **Read the rule** shows which one.
- **This assistant can reach a service you have not connected**: the assistant's tools act with your own account. For a service you connect with a token, click **Connect my account** and paste it; the conversation then continues with the tools available. For a service that signs you in on its own page, click **Connect in Integrations**, connect there, then come back and send your message again. See [Integrations](https://agentlab.cresclab.com/docs/en/studio/integrations.md#authorize-in-chat).

For tools and accounts, see [Integrations](https://agentlab.cresclab.com/docs/en/studio/integrations.md).

## The conversation list

Above the list, switch between three views:

| View          | What it shows                                                                          |
| ------------- | -------------------------------------------------------------------------------------- |
| **Recent**    | Conversations you started, newest first, with your pins in a **Pinned** section on top |
| **Put away**  | Conversations you put away                                                             |
| **Automatic** | Conversations your automations produced, with the schedule that ran and how it ended   |

The list loads in pages; **Show older conversations** loads more.

Automations are work an assistant runs on a clock. See [Automations](https://agentlab.cresclab.com/docs/en/studio/automations.md).

## New-result marks

An automation leaves its result in a new conversation, often while you are away from the console.

- A dot next to a conversation means **New since you last opened it**.
- The **Automatic** button and the **Conversations** entry in the left rail show how many results you haven't opened, up to "99+".
- Opening the conversation counts as seeing it, and the mark goes away.

Marks are about your own conversations only. An admin opening your conversation doesn't clear your mark.

## Search conversations

The search box above the list finds conversations by what was said in them. Press Enter or the magnifier.

- A conversation matches only when it contains every word you typed.
- Each result shows the passage that matched.
- Put-away conversations are searched too.
- Search covers your 200 most recent conversations.

## Organize a conversation

Open a conversation and use the menu at the top right:

| Action                        | What it does                                                                                    |
| ----------------------------- | ----------------------------------------------------------------------------------------------- |
| **Rename**                    | Up to 80 characters. **Use the original name** restores the generated one.                      |
| **Pin** / **Unpin**           | Up to 20 pinned conversations. A put-away conversation can't be pinned.                         |
| **Put away** / **Bring back** | Moves it to **Put away** and removes its pin. Bringing it back doesn't restore the pin.         |
| **Conversation details**      | Shows the assistant, the conversation's **Tasks** checklist, and **File export and retention**. |

Only the person who started a conversation can do these. The console has no way to delete a conversation; put away the ones you don't want to see.

**Tasks** in **Conversation details** is a checklist for that one conversation. Add an item with **Add**, then mark it **Start**, **Done** or **Skip**. It is separate from the **Tasks** page in the left rail; for that, see [Tasks](https://agentlab.cresclab.com/docs/en/studio/tasks.md).

After a reply, **Save as a test question** under the message turns that exchange into a test question for the assistant, checked before every future publish.

## Who can see a conversation

Conversations can't be shared.

- Only you see the conversations you started, in the list and in search.
- Workspace owners and admins can read any member's conversation. Each such read and search is recorded in the audit trail.
- A conversation that started in Slack is read-only here; continue it in Slack. You can still approve its pending actions here.