> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qobra.co/llms.txt
> Use this file to discover all available pages before exploring further.

# get_ai_conversation

> Read one AI conversation page by page: its summary plus the messages exchanged with the agent and the tools it ran for each answer.

Return one AI conversation: its
[`list_ai_conversations`](https://docs.qobra.co/mcp_documentation/tools/list_ai_conversations) row
plus a page of its transcript, oldest message first.

Tool calls are not messages of their own. Each answer names, in `tools`, the
tools the agent ran before it, up to 8, and counts the rest in
`other_tool_count`. Each message is cut at 1,500 characters: `truncated` says
so, and the audit trail's **AI conversations** tab shows it whole. Messages
undone by an Architect rollback stay in the transcript, marked `reverted`.

The tool reads the conversation as stored and does not wait for a turn in
progress. A transcript that ends on the author's message, or on an answer with
no text, is a turn still running, waiting on its author, or one that ended
without answering.

Keep paginating while `next_offset` is set.

Requires read access to the audit trail.

```json Example arguments theme={null}
{
  "conversation_id": "665f1c2e8a4b3d0012a7e9f1",
  "limit": 40
}
```

## Parameters

> [accordion] Body
>
> **conversation\_id** `string` *required*
>
> Id from
> [`list_ai_conversations`](https://docs.qobra.co/mcp_documentation/tools/list_ai_conversations),
> or from an audit log's `conversation`.
>
> **limit** `integer`
>
> Messages per page, 1 to 40. Default 15.
>
> **offset** `integer`
>
> Messages to skip. Omit it for the first page; for the next one pass the
> previous next\_offset.

## Response

> [accordion] Body
>
> **id** `string`
>
> **resource\_url** `string`
>
> Link opening this conversation in the audit trail.
>
> **agent\_type** `string`
>
> The agent talked to: 'architect' (designs compensation plans), 'analyst'
> (builds reports, flagging records included), 'qobra\_ai' (the assistant every
> user talks to), 'facets' (builds Facets), 'sales\_coach' (answers questions
> on a statement). One of: `architect`, `analyst`, `qobra_ai`, `facets`,
> `sales_coach`.
>
> **title** `string`
>
> The prompt the conversation was opened with, cut at 200 characters
> (`title_truncated` says when). Null on a conversation opened without one.
>
> **title\_truncated** `boolean`
>
> **author** `object`
>
> The user who talked to the agent. Null when they are no longer a user of the
> company.
>
> **object properties**
>
> **id** `string`
>
> **email** `string`
>
> **name** `string`
>
> **started\_at** `string`
>
> **last\_active\_at** `string`
>
> **message\_count** `integer`
>
> Messages the transcript holds: the author's, the agent's answers and the
> app's notices, its tool calls folded into the answers.
>
> **write\_count** `integer`
>
> Changes the conversation saved, as audit-trail entries: list\_audit\_logs with
> this `conversation_id` lists them, beside the entry of its opening. A
> compensation edit not saved yet is not counted.
>
> **messages** `object[]`
>
> A page of the transcript, oldest message first.
>
> **object properties**
>
> **role** `string`
>
> 'user' (the author), 'assistant' (the agent's answer) or 'system' (a
> notice of the app). Tool calls are in the answer's `tools`, not messages
> of their own. One of: `user`, `assistant`, `system`, `tool`.
>
> **content** `string`
>
> The text, cut at 1,500 characters (`truncated` says when). The audit
> trail's AI conversations tab shows it whole.
>
> **truncated** `boolean`
>
> **reverted** `boolean`
>
> True on a message the user rolled back in the Architect: its draft went
> back to before the first message so marked, and the conversation went on
> from there.
>
> **timestamp** `string`
>
> **tools** `string[]`
>
> Tools the agent ran before this message, each once, in the order first
> called, the first 8 of them: on an answer, or on the notice of the app a
> failed turn ended on. Empty on a user message and when it ran none.
>
> **other\_tool\_count** `integer`
>
> Further tools it ran, not named in `tools`.
>
> **files** `string[]`
>
> Names of the files sent with the message.
>
> **next\_offset** `integer`
>
> Offset of the transcript's next page; null on its last.
